Vous êtes sur la page 1sur 34

J Math Imaging Vis

DOI 10.1007/s10851-012-0378-3

Scale Selection Properties of Generalized Scale-Space Interest


Point Detectors
Tony Lindeberg

© The Author(s) 2012. This article is published with open access at Springerlink.com

Abstract Scale-invariant interest points have found several scale normalized derivatives for each one of the pure second-
highly successful applications in computer vision, in partic- order operators. In this respect, no scale compensation is
ular for image-based matching and recognition. needed between the two types of scale selection approaches.
This paper presents a theoretical analysis of the scale se- When using post-smoothing, the scale estimates may, how-
lection properties of a generalized framework for detecting ever, be different between different types of interest point
interest points from scale-space features presented in Linde- operators, and it is shown how relative calibration factors
berg (Int. J. Comput. Vis. 2010, under revision) and com- can be derived to enable comparable scale estimates for each
prising: purely second-order operator and for different amounts of
self-similar post-smoothing.
– an enriched set of differential interest operators at a fixed
A theoretical analysis of the sensitivity to affine image
scale including the Laplacian operator, the determinant of
deformations is presented, and it is shown that the scale esti-
the Hessian, the new Hessian feature strength measures I
mates obtained from the determinant of the Hessian operator
and II and the rescaled level curve curvature operator, as
are affine covariant for an anisotropic Gaussian blob model.
well as
Among the other purely second-order operators, the Hessian
– an enriched set of scale selection mechanisms including
feature strength measure I has the lowest sensitivity to non-
scale selection based on local extrema over scale, com-
uniform scaling transformations, followed by the Laplacian
plementary post-smoothing after the computation of non-
operator and the Hessian feature strength measure II. The
linear differential invariants and scale selection based on
predictions from this theoretical analysis agree with experi-
weighted averaging of scale values along feature trajecto-
mental results of the repeatability properties of the different
ries over scale.
interest point detectors under affine and perspective trans-
It is shown how the selected scales of different linear and formations of real image data. A number of less complete
non-linear interest point detectors can be analyzed for Gaus- results are derived for the level curve curvature operator.
sian blob models. Specifically it is shown that for a rota-
tionally symmetric Gaussian blob model, the scale estimates Keywords Feature detection · Interest point · Blob
obtained by weighted scale selection will be similar to the detection · Corner detection · Scale · Scale-space · Scale
scale estimates obtained from local extrema over scale of selection · Scale invariance · Scale calibration · Scale
linking · Feature trajectory · Deep structure · Affine
transformation · Differential invariant · Gaussian
The support from the Swedish Research Council, Vetenskapsrådet derivative · Multi-scale representation · Computer vision
(contracts 2004-4680, 2010-4766) and from the Royal Swedish
Academy of Sciences as well as the Knut and Alice Wallenberg
Foundation is gratefully acknowledged.
1 Introduction
T. Lindeberg ()
School of Computer Science and Communication, KTH Royal
Institute of Technology, 100 44 Stockholm, Sweden The notion of scale selection is essential to adapt the scale
e-mail: tony@csc.kth.se of processing to local image structures. A computer vision
J Math Imaging Vis

system equipped with an automatic scale selection mecha- The subject of this article is to perform an in-depth theoret-
nism will have the ability to compute scale-invariant image ical analysis of properties of these scale selection methods
features and thereby handle the a priori unknown scale vari- when applied to the task of computing scale-invariant inter-
ations that may occur in image data because of objects and est points:
substructures of different physical size in the world as well
(i) When using a set of different types of interest point de-
as objects at different distances to the camera. Computing
tectors that are based on different linear or non-linear
local image descriptors at integration scales proportional to
combinations of scale-space derivatives, a basic ques-
the detection scales of scale-invariant image features, more-
tion arises of how to relate thresholds on the magnitude
over makes it possible to compute scale-invariant image
descriptors (Lindeberg [35]; Bretzner and Lindeberg [4]; values between different types of interest point detec-
Mikolajczyk and Schmid [49]; Lowe [48]; Bay et al. [2]; tors. By studying the responses of the different inter-
Lindeberg [38, 43]). est point detectors to unit contrast Gaussian blobs, we
A general framework for performing scale selection can will derive a way of expressing mutually correspond-
be obtained by detecting local extrema over scale of γ - ing thresholds between different types of interest points
normalized derivative expressions (Lindeberg [35]). This detectors. Algorithmically, the resulting threshold rela-
approach has been applied to a large variety of feature detec- tions lead to intuitively very reasonable results.
tion tasks (Lindeberg [34]; Bretzner and Lindeberg [4]; Sato (ii) The new scale selection method based on weighted
et al. [54]; Frangi et al. [11]; Krissian et al. [22]; Chomat averaging along feature trajectories over scale raises
et al. [5]; Hall et al. [15]; Mikolajczyk and Schmid [49]; questions of how the properties of this scale selection
Lazebnik et al. [24]; Negre et al. [52]; Tuytelaars and Miko- method can be related to the previous scale selection
lajczyk [58]). Specifically, highly successful applications method based on local extrema over scale of scale-
can be found in image-based recognition (Lowe [48]; Bay normalized derivatives. We will show that for Gaussian
et al. [2]). Alternative approaches for scale selection have blobs, the scale estimates obtained by weighted aver-
also been proposed in terms of the detection of peaks over aging over scale will be similar to the scale estimates
scale in weighted entropy measures (Kadir and Brady [18]) obtained from local extrema over scale. If we assume
or Lyapunov functionals (Sporring et al. [56]), minimization that scale calibration can be performed based on the be-
of normalized error measures over scale (Lindeberg [36]), haviour for Gaussian blobs, this result therefore shows
determining minimum reliable scales for feature detection that no relative scale compensation is needed between
according a noise suppression model (Elder and Zucker [9]), the two types of scale selection approaches. In previous
determining optimal stopping times in non-linear diffusion- work on scale selection based on γ -normalized deriva-
based image restoration methods using similarity measure- tives [34, 35] a similar assumption of scale calibra-
ments relative to the original data (Mrázek and Navara [51]), tion based on Gaussian model signals has been demon-
by applying statistical classifiers for texture analysis at dif- strated to lead to highly useful results for calibrating
ferent scales (Kang et al. [19]) or by performing image the value of the γ -parameter with respect to the prob-
segmentation from the scales at which a supervised clas- lems of blob detection, corner detection, edge detection
sifier delivers class labels with the highest posterior (Loog and ridge detection, with a large number of successful
et al. [47]; Li et al. [25]). computer vision applications building on the resulting
Recently, a generalization of the differential approach feature detectors.
for scale selection based on local extrema over scale of γ - (iii) For the scale linking algorithm presented in [39],
normalized derivatives has been proposed by linking image which is based on local gradient ascent or gradient de-
features over scale into feature trajectories over scale in cent starting from local extrema in the differential re-
a generalized scale-space primal sketch [39]. Specifically, sponses at adjacent levels of scale, it turns out that a
two novel scale selection mechanisms have been proposed second post-smoothing stage after the computation of
in terms of: non-linear differential invariants is highly useful for
– post-smoothing of differential feature responses by per- increasing the performance of the scale linking algo-
forming a second-stage scale-space smoothing step after rithm, by suppressing spurious responses of low rela-
the computation of non-linear differential invariants, so as tive amplitude in the non-linear differential responses
to simplify the task of linking feature responses over scale that are used for computing interest points. This self-
into feature trajectories, and similar amount of post-smoothing is determined as a
– weighted scale selection where the scale estimates are constant times the local scale for computing the differ-
computed by weighted averaging of scale-normalized fea- ential expressions, and may affect the scale estimates
ture responses along each feature trajectory over scale, in obtained from local extrema over scale or weighted av-
contrast to previous detection of local extrema or global eraging over scale. We will analyze how large this ef-
extrema over scale. fect will be for different amounts of post-smoothing
J Math Imaging Vis

and also show how relative scale normalization fac- In very recent work [42], these generalized scale-space
tors can be determined for the different differential ex- interest points have been integrated with local scale-invariant
pressions to obtain scale estimates that are unbiased image descriptors and been demonstrated to lead to highly
with respect to the effect of the post-smoothing op- competitive results for image-based matching and recogni-
eration, if we again assume that scale calibration can tion.
be performed based on the scale selection properties
for Gaussian blobs. Notably, different scale compen- 1.1 Outline of the Presentation
sation factors for the influence of post-smoothing will
be obtained for the different differential expressions The paper is organized as follows. Section 2 reviews main
that are used for defining interest points. Without post- components of a generalized framework for detecting scale-
smoothing, the scale estimates obtained from the dif- invariant interest points from scale-space features, including
ferent differential expressions are, however, all similar a richer set of interest point detectors at a fixed scale as well
for Gaussian blobs, which indicates the possibilities of as new scale selection mechanisms.
using different types of differential expressions for per- In Sect. 3 the scale selection properties of this framework
forming combined interest point detection and scale se- are analyzed for scale selection based on local extrema over
lection, so that they can be interchangeably replaced in scale of γ -normalized derivatives, when applied to rotation-
a modular fashion. ally symmetric as well as anisotropic Gaussian blob models.
(iv) When detecting interest points from images that are Section 4 gives a corresponding analysis for scale selection
taken of an object from different viewing directions, by weighted averaging over scale along feature trajectories.
the local image pattern will be deformed by the per- Section 5 summarizes and compares the results obtained
spective projection. If the interest point corresponds to from the two scale selection approaches including comple-
a point in the world that is located at a smooth surface mentary theoretical arguments to highlight their similarities
of an object, this deformation can to first order of ap- in the rotationally symmetric case. It is also shown how scale
proximation be modelled by a local affine transforma- calibration factors can be determined so as to obtain compa-
tion (Gårding and Lindeberg [12]). While the notion rable scale estimates from interest point detectors that have
of affine shape adaptation has been demonstrated to be been computed from different types of differential expres-
a highly useful tool for computing affine invariant in- sions. Comparisons are also presented of the relative sensi-
terest points (Lindeberg and Gårding [46]; Baumberg tivity of the scale estimates to affine transformations outside
[1]; Mikolajczyk and Schmid [49]; Tuytelaars and van the similarity group, with a brief comparison to experimen-
Gool [57]), the success of such an affine shape adapta- tal results. Finally, Sect. 6 concludes with an overall sum-
tion process depends on the robustness of the underly- mary and discussion.
ing interest points that are used for initiating the itera-
tive affine shape adaptation process. To investigate the
properties of the different interest point detectors under 2 Scale-Space Interest Points
affine transformations, we will perform a detailed anal-
ysis of the scale selection properties for affine Gaussian 2.1 Scale-Space Representation
blobs, for which closed form theoretical analysis is pos-
sible. The analysis shows that the determinant of the The conceptual background we consider for feature detec-
Hessian operator and the new Hessian feature strength tion is a scale-space representation (Iijima [17]; Witkin
measure I do both have significantly better behaviour [62]; Koenderink [20]; Koenderink and van Doorn [21]; Lin-
under affine transformations than the Laplacian opera- deberg [30, 31]; Florack [10]; Weickert et al. [60]; ter Haar
tor or the new Hessian feature strength measure II. In Romeny [14]; Lindeberg [38, 40]) L : R2 × R+ → R com-
comparison with experimental results [39], the interest puted from a two-dimensional signal f : R2 → R according
point detectors that have the best theoretical properties to
under affine transformations of Gaussian blob do also 
have significantly better repeatability properties under L(x, y; t) = f (x − u, y − v) g(u, v; t) du dv (1)
affine and perspective transformations than the other (u,v)∈R2

two. These results therefore show how experimental


where g : R2 × R+ → R denotes the (rotationally symmet-
properties of interest points can be predicted by the-
ric) Gaussian kernel
oretical analysis, which contributes to an increased un-
derstanding of the relative properties of different types 1 −(x 2 +y 2 )/2t
g(x, y; t) = e (2)
of interest point detectors. 2πt
J Math Imaging Vis

and the variance t = σ 2 of this kernel is referred to as the – the determinant of the Hessian
scale parameter. Equivalently, this scale-space family can
be obtained as the solution of the (linear) diffusion equation det HL = Lxx Lyy − L2xy (6)

1 – the rescaled level curve curvature


∂t L = ∇ 2 L (3)
2
κ̃(L) = L2x Lyy + L2y Lxx − 2Lx Ly Lxy (7)
with initial condition L(·, ·; t) = f . From this representa-
tion, scale-space derivatives or Gaussian derivatives at any (ii) either of the following new differential analogues and
scale t can be computed either by differentiating the scale- extensions of the Harris operator [16] proposed in [39]:
space representation or by convolving the original image – the unsigned Hessian feature strength measure I
with Gaussian derivative kernels: ⎧
⎨ det HL − k trace HL
2

 
Lx α y β (·, ·; t) = ∂x α y β L(·, ·; t) = ∂x α y β g(·, ·; t) ∗ f (·, ·) D1 L = if det HL − k trace2 HL > 0 (8)


(4) 0 otherwise

– the signed Hessian feature strength measure I


where α and β ∈ Z+ .


⎪ det HL − k trace2 HL


2.2 Differential Entities for Detecting Scale-Space Interest ⎪

⎨ if det HL − k trace HL > 0
2
Points
D̃1 L = det HL + k trace2 HL (9)



⎪ if det HL + k trace2 HL < 0
A common approach to image matching and object recog- ⎪


nition consists of matching interest points with associated 0 otherwise
image descriptors. Basic requirements on the interest points
on which the image matching is to be performed are that where k ∈]0, 14 [ with the preferred choice k ≈ 0.04, or
they should (i) have a clear, preferably mathematically well- (iii) either of the following new differential analogues and
founded, definition, (ii) have a well-defined position in im- extensions of the Shi and Tomasi operator [55] pro-
age space, (iii) have local image structures around the inter- posed in [39]:
est point that are rich in information content such that the – the unsigned Hessian feature strength measure II
interest points carry important information to later stages
and (iv) be stable under local and global deformations of D2 L = min(|λ1 |, |λ2 |) = min(|Lpp |, |Lqq |) (10)
the image domain, including perspective image deforma-
tions and illumination variations such that the interest points – the signed Hessian feature strength measure II
can be reliably computed with a high degree of repeatability. ⎧

⎨ Lpp if |Lpp | < |Lqq |
The image descriptors computed at the interest points should
D̃2 L = Lqq if |Lqq | < |Lpp | (11)
also (v) be sufficiently distinct, such that interest points cor- ⎪

responding to physically different points can be kept sepa- (Lpp + Lqq )/2 otherwise
rate.
Preferably, the interest points should also have an at- where Lpp and Lqq denote the eigenvalues of the Hes-
sian matrix (the principal curvatures) ordered such that
tribute of scale, to make it possible to compute reliable in-
Lpp ≤ Lqq [34]:
terest points from real-world image data, including scale
changes in the image domain. Specifically, the interest 1

points should preferably also be scale-invariant to make it Lpp = Lxx + Lyy − (Lxx − Lyy )2 + 4L2xy (12)
2
possible to match corresponding image patches under scale
1

variations. Lqq = Lxx + Lyy + (Lxx − Lyy )2 + 4L2xy (13)
2
Within this scale-space framework, interest point detec-
tors can be defined at any level of scale using Figure 1 shows examples of detecting different types of in-
terest points from a grey-level image. In this figure, the
(i) either of the following established differential opera-
repetitive nature of the underlying image structures in the
tors [35]:
row of similar books illustrate the ability of the interest point
– the Laplacian operator detectors to respond to approximately similar structures in
the image domain by corresponding responses. Figure 2 il-
∇ 2 L = Lxx + Lyy (5) lustrates the repeatability properties of such interest points
J Math Imaging Vis

Fig. 1 Scale-invariant interest points detected by linking (top left) scale selection by weighted averaging of scale values along each fea-
Laplacian ∇norm
2 L features, (top right) determinant of the Hessian ture trajectory. The 500 strongest interest points have been extracted
det Hnorm L features, (middle left) signed Hessian feature strength and drawn as circles with
√ the radius reflecting the selected scale mea-
measure D̃1,norm L features, (middle right) signed Hessian feature sured in units of σ = t . Positive responses of the differential expres-
strength measure D̃2,norm L features, (bottom left) rescaled level curve sion DL are shown in red and negative responses in blue. (Image size:
curvature κ̃γ −norm (L) features and (bottom right) scale-linked Harris- 725 × 480 pixels. Scale range: t ∈ [4, 512])
Laplace features over scale into feature trajectories and performing
J Math Imaging Vis


Fig. 2 Illustration of the repeatability properties of the interest points scale measured in units of σ = t . Interest points that have a posi-
by detecting signed Hessian feature strength D̃1,norm L interest points tive definite Hessian matrix are shown in blue (dark features), interest
from two images of a building taken from different perspective views, points with negative definite Hessian matrix are shown in red (bright
by linking image features over scale into feature trajectories and per- features) whereas interest points with an indefinite Hessian matrix are
forming scale selection by weighted averaging of scale values along marked in green (saddle-like features). (Image size: 816 × 540 pixels.
each feature trajectory. The 1000 strongest interest points have been Scale range: t ∈ [4, 256])
extracted and drawn as circles with the radius reflecting the selected
J Math Imaging Vis

more explicitly, by detecting signed Hessian feature strength Other ways of defining image features from the second-
D̃1,norm L interest points from two images of a building taken order differential image structure of images have been pro-
from different perspective views. posed by Danielsson et al. [7] and Griffin [13].
A basic motivation for defining the new differential op-
erators D1 , D̃1 , D2 and D̃2 from the Hessian matrix HL
2.3 Scale Selection Mechanisms
in a structurally related way as the Harris and the Shi-and-
Tomasi operators are defined from the second-moment ma-
trix (structure tensor) are that: (i) under an affine trans- Scale Selection from γ -Normalized Derivatives In (Linde-
formation p = A p with p = (x, y)T and A denoting a berg [29, 31, 35, 37]) a general framework for automatic
non-singular 2 × 2 matrix it can be shown that the Hes- scale selection was proposed based on the idea of detect-
sian matrix Hf transforms in a similar way (Hf )(p ) = ing local extrema over scale of γ -normalized derivatives de-
A−T (Hf )(p) A−1 as the second-moment matrix μ (p) = fined according to
A−T μ(p) A−1 [31, 46] and (ii) provided that the Hessian
matrix is either positive or negative definite, the Hessian ∂ξ = t γ /2 ∂x , ∂η = t γ /2 ∂y (15)
matrix HL computed at a point p0 defines an either pos-
itive or negative definite quadratic form QHL (p) = (p − where γ > 0 is a free parameter1 that can be related to the
p0 )T (HL)(p − p0 ) in a similar way as the second-moment dimensionality of the image features that the feature detec-
matrix μ computed at p0 does: Qμ (p) = (p − p0 )T μ (p − tor is designed to respond to, e.g., in terms of the evolution
p0 ). From these two analogies, we can conclude that pro- properties over scale in terms of (i) self-similar Lp -norms of
vided the Hessian matrix is either positive or negative defi-
Gaussian derivative operators for different dimensionalities
nite, these two types of descriptors should have strong quali-
of the image space [35, Sect. 9.1], (ii) self-similar Fourier
tative similarities. Experimentally, the new differential inter-
spectra [35, Sect. 9.2] or (iii) the fractal dimension of the
est point detectors D1 , D̃1 , D2 and D̃2 can be shown to per-
image data [53]; see also Appendix A.3 for an explicit inter-
form very well and to allow for image features with better
repeatability properties under affine and perspective trans- pretation of the parameter γ in terms of the dimensionality
formations than the more traditional Laplacian or Harris op- D of second-order image features according to (213).
erators [39]. Specifically, it was shown in [35] that local extrema
The Laplacian ∇ 2 L responds to bright and dark blobs over scale of homogeneous polynomial differential invari-
as formalized in terms of local minima or maxima of the ants Dγ −norm L expressed in terms of γ -normalized Gaus-
Laplacian operator. The determinant of the Hessian det HL sian derivatives are transformed in a scale-covariant way:
responds to bright and dark blobs by positive responses and
If some scale-normalized differential invariant
in addition to saddle-like image features by negative re-
Dγ −norm L assumes a local extremum over scale at
sponses as well as to corners. The unsigned Hessian feature
strength D1 L responds to bright and dark blobs as well as scale t0 in scale-space, then under a uniform rescal-
to corners, with the complementary requirement that the ra- ing of the input pattern by a factor s there will be a
tio of the eigenvalues λ1 and λ2 of the Hessian matrix (with local extremum over scale in the scale-space of the
|λ1 | ≤ |λ2 ) should be sufficiently close to one, as specified transformed signal at scale s 2 t0 .
by the parameter k according to:
Furthermore, by performing simultaneous scale selection
2k |λ1 | and spatial selection by detecting scale-space extrema,
√ ≤ ≤1 (14)
1 − 2k + 1 − 4k |λ2 | where the scale-normalized differential expression
Dγ −norm L assumes local extrema with respect to both space
For this entity to respond, it is therefore necessary that there
and scale, constitutes a general framework for detecting
are strong intensity variations along two different directions
scale-invariant interest points. Formally, such scale-space
in the image domain. The signed Hessian feature strength
extrema are characterized by the first-order derivatives with
measure D̃1 L responds to similar image features as the un-
signed entity D1 L, and in addition to saddle-like image fea- respect to space and scale being zero
tures with a corresponding constraint on the ratio between
the eigenvalues. The Hessian feature strength measures D2 L ∇(Dγ −norm L) = 0 and ∂t (Dγ −norm L) = 0 (16)
and D̃2 L respond strongly when both of the principal cur-
vatures are strong and the local image pattern therefore con- 1 Indeed, it can be shown that the definition of scale-normalized deriva-
tains strong intensity variations in two orthogonal directions.
tives in this way captures the full degrees of freedom by which scale
The unsigned entity D2 L disregards the sign of the princi- invariance can be obtained from local extrema over scale of scale-
pal curvatures, whereas the signed entity D̃2 L preserves the normalized derivatives defined from a Gaussian scale-space, as for-
sign of the principal curvature of the lowest magnitude. mally proved by necessity in [35, Appendix A.1].
J Math Imaging Vis

and in addition the composed Hessian matrix computed over local extrema over scale in the sense that interesting charac-
both space and scale teristic scale levels for further analysis should be obtained
from the scales at which the differential operator assumes
⎛ ⎞ is strongest scale-normalized magnitude values over scale.
∂xx ∂xy ∂xt

H(x,y; t) (Dγ −norm L) = ∂xy ∂yy ∂yt ⎠ (Dγ −norm L) Contrary to scale selection based on local extrema over
∂xt ∂yt ∂tt scale, however, scale selection by weighted averaging over
scale implies that the scale estimate will not only be obtained
(17) from the behaviour around the local extremum over scale,
but also including the responses from all scales along a fea-
being either positive or negative definite. ture trajectory over scale. The intention behind this choice is
that the scale estimates should therefore be more robust and
Generalized Scale Selection Mechanisms In [39] this ap- less sensitive to local image perturbations.
proach was extended in the following ways: Experimentally, it can be shown that scale-space interest
– by performing post-smoothing of the differential expres- points detected by these generalized scale selection mecha-
sion Dγ −norm L prior to the detection of local extrema nisms lead to interest points with better repeatability proper-
over space or scale ties under affine and perspective image deformations com-
pared to corresponding interest points detected by regular
(Dγ −norm L)(x, y; t) scale-space extrema [39]. In this sense, these generalized
 scale selection mechanisms make it possible to detect more
= (Dγ −norm L)(x − u, y − v; t) robust image features. Specifically, the use of scale selec-
(u,v)∈R2 tion by weighted averaging over scale is made possible by
  linking image features over scale into feature trajectories,2
× g u, v; c2 t du dv (18)
which ensures that the scale estimates should only be in-
with an integration scale (post-smoothing scale) tpost = fluenced by responses from scale levels that correspond to
c2 t proportional to the differentiation scale t with c > 0 qualitatively similar types of image structures along a fea-
(see Appendix A.1 for a brief description of the algorith- ture trajectory over scale.
mic motivations for using such a post-smoothing opera- The subject of this article is to analyze properties of these
tion when linking image features over scale that have been generalized scale selection mechanisms theoretically when
computed from non-linear differential entities) and applied to the interest point detectors listed in Sect. 2.2.
– by performing weighted averaging of scale values along
any feature trajectory T over scale in a scale-space primal
3 Scale Selection Properties for Local Extrema over
sketch according to
Scale

τ ψ((Dγ −norm L)(x(τ ); τ )) dτ
τ̂T = τ ∈T (19) For theoretical analysis, we will consider a Gaussian pro-
τ ∈T ψ((Dγ −norm L)(x(τ ); τ )) dτ totype model of blob-like image structures. With such a
prototype model, the semi-group property of the Gaussian
where ψ denotes some (positive and monotonically in-
kernel makes it possible to directly obtain the scale-space
creasing) transformation of the scale-normalized feature
representations at coarser scales in terms of Gaussian func-
strength response Dγ −norm L and with the scale parame-
tions, which simplifies theoretical analysis. Specifically, the
ter parameterized in terms of effective scale [28]
result of computing polynomial differential invariants at dif-
ferent scales will be expressed in terms of Gaussian func-
τ = A log t + B where A ∈ R+ and B ∈ R (20)
tions multiplied by polynomials. Thereby, closed-form the-
to obtain a scale covariant construction of the correspond- oretical analysis becomes tractable, which would otherwise
ing scale estimates
  2 By linking image features over scale into feature trajectories it also
τ̂T − B becomes possible to define a significance value by integrating scale-
tˆT = exp (21) normalized feature responses over scale. Experimentally, it can be
A
shown that such ranking of image features leads to selections of sub-
that implies that the resulting image features will be scale- sets of interest points with better overall repeatability properties than
selection of subsets of interest points from the extremum responses of
invariant. interest points detectors at scale-space extrema. An intuitive motiva-
tion for this property is a heuristic principle that image features that
The motivation for performing scale selection by weighted are stable over large ranges of scales should be more likely to be sig-
averaging of scale-normalized differential responses over nificant than image features that only exist over a shorter life length in
scale is analogous to the motivation for scale selection from scale-space [27, Assumption 1 in Sect. 3 on p. 296].
J Math Imaging Vis

be much harder to carry out regarding the application of the derivatives over scale to a single Gaussian blob:
non-linear operations that are used for defining the interest
points to general image data. 1 − x 22t+y 2
f (x, y) = g(x, y; t0 ) = e 0 (22)
The use of Gaussian prototype model can also be moti- 2πt0
vated by conceptual simplicity. If we would like to model
Due to the semi-group property of the Gaussian kernel
an image feature at some scale, then the Gaussian model is
the model that requires the minimum amount of information
g(·, ·; t1 ) ∗ g(·, ·; t2 ) = g(·, ·; t1 + t2 ) (23)
in the sense that the Gaussian distribution is the distribution
with maximum entropy3 given a specification of the mean the scale-space representation of f obtained by Gaussian
value m and the covariance matrix Σ of the distribution. smoothing is given by
Specifically, the Gaussian function with scale parameter t
2 2
serves as an aperture function that measures image struc- 1 x +y
− 2(t
tures with respect to an inner scale beyond which finer-scale L(x, y; t) = g(x, y; t0 + t) = e 0 +t)
(24)
2π(t0 + t)
structures cannot be resolved.
In previous work [34, 35] it has been shown that deter- 3.1.1 The Pure Second-Order Interest Point Detectors
mination of the γ -parameter in scale selection for different
types of feature detection tasks, such as blob detection, cor- By differentiation, if follows that the scale normalized
ner detection, edge detection and ridge detection, can be per- (signed or unsigned) feature strength measure at the cen-
formed based on the behaviour of these feature detectors on ter (x, y) = (0, 0) of the blob will for the Laplacian (5),
Gaussian-based intensity profiles. As will be shown later, the determinant of the Hessian (6) and the Hessian feature
the theoretical results that will be derived based on Gaussian strength measures I (8) and II (10) be given by
blob models will lead to theoretical predictions that agree
with the relative repeatability properties of different types  2  tγ
∇γ −norm L (0, 0; t) = − (25)
of interest point detectors under affine and perspective trans- π(t0 + t)2
formations. Formally, however, further application of these
t 2γ
results will be based on an assumption that the scale selec- (det Hγ −norm L)(0, 0; t) = (26)
tion behaviour can be calibrated based on the behaviour for 4π 2 (t0 + t)4
Gaussian prototype models. (1 − 4k) t 2γ
(D1,γ −norm L)(0, 0; t) = (27)
4π 2 (t0 + t)4
3.1 Regular Scale Selection from Local Extrema over

Scale (D2,γ −norm L)(0, 0; t) = (28)
2π(t0 + t)2
Two basic questions in the relation to the different interest
By differentiating these expressions with respect to the scale
point detectors reviewed in Sect. 2.2 concern:
parameter t and setting the derivative to zero, it follows that
– How will the selected scale levels be related between dif- the extremum value over scale will for all these descriptors
ferent interest point detectors? be assumed at the same scale
– How will the scale-normalized magnitude values be re- γ
lated between different interest point detectors that re- tˆ = t0 (29)
2−γ
spond to similar image structures?
Ideally, we would like similar scale estimates to be obtained For the specific choice of γ = 1, the selected scale tˆ will
for different interest point detectors, so that the interest point be equal to the scale of the Gaussian blob, i.e. tˆ = t0 , and
detectors could be modularly replaceable in the computer vi- the extremum value over scale for each one of the respective
sion algorithms they are part of. Since the interest point de- feature detectors is
tectors are expressed in terms of different types of linear or  2  1
∇  = (30)
non-linear combinations of scale-space derivatives, a basic norm L max
4πt0
question concerns how to express comparable thresholds on
the magnitude values for the different interest point detec- 1
(det Hnorm L)max = (31)
tors. In this section, we will relate these entities by apply- 64π 2 t02
ing scale selection from local extrema of scale-normalized (1 − 4k)
(D1,norm L)max = (32)
64π 2 t02
3 Maximum entropy solutions have been argued to be taken as preferred

default solutions for underconstrained problems [3, 59] although the 1


(D2,norm L)max = (33)
applicability of these arguments has also been questioned [6, 8]. 8πt0
J Math Imaging Vis

Table 1 Relationships between


scale-normalized thresholds Feature detector DL CD L
CDL for different types of
scale-invariant interest point Laplacian ∇ 2 Lnorm = t (Lxx + Lyy ) C∇ 2 L = C
detectors DL = ∇ 2 L, det HL, determinant of the Hessian det Hnorm L = t 2 (L xx Lyy − Lxy )
2 Cdet HL = C 2 /4
D1 L, D̃1 L, D2 L and D̃2 L using
Hessian feature strength I D1,norm L = t (Lxx Lyy − Lxy − k (Lxx
2 2 + Lyy )2 ) CD1 L = (1 − 4k) C 2 /4
scale-normalized derivatives
with γ = 1. The complementary Hessian feature strength Ĩ D̃1,norm L = t 2 (Lxx Lyy − L2xy ± k (Lxx + Lyy )2 ) CD̃1 L = (1 − 4k) C 2 /4
expression for the Hessian feature strength II D2,norm = t min(|Lpp |, |Lqq |) CD2 L = C/2
Harris-Laplace operator is based ˜
Hessian feature strength II D̃2,norm L = t (Lpp or Lqq ) CD̃2 L = C/2
on the assumption of a relative
integration scale of r = 1 Harris-Laplace Hnorm = t 2 (det μ − k trace2 μ) CH = (1 − 4k) C 4 /256

These results are in full agreement with earlier results about Note: For the Harris operator [16], which is determined
the scale selection properties for Gaussian blobs concerning from the second-moment matrix
the scale-normalized Laplacian and the scale-normalized
μ(x, y; t, s)
determinant of the Hessian [31, Sect. 13.3.1] [35, Sect. 5.1].   
L2x (u, v; t) Lx (u, v; t) Ly (u, v; t)
=
(u,v)∈R2 Lx (u, v; t) Ly (u, v; t) L2y (u, v; t)
3.1.2 Scale Invariant Feature Responses After Contrast
× g(x − u, y − v; s) du dv (38)
Normalization
according to
When applying different types of interest point detectors in  
Hnorm = t 2 det μ − k trace2 μ (39)
parallel, some approach is needed for expressing compara-
ble thresholds between different types of interest point de- for some k ∈]0, 14 [, a corresponding analysis shows that the
tectors. Let us assume that such calibration of corresponding response at the center (x, y) = (0, 0) of a Gaussian blob is
thresholds between different interest point detectors can be at scale t = t0 given by
performed based on the their responses to Gaussian blobs.
If we would like to present a Gaussian blob on a screen and (1 − 4k) r 4
Hmax = (40)
would like to make it possible to vary its size (spatial extent) 256(r 2 + 1)4
without affecting its perceived brightness on the screen, let
if we let the integration scale s be related to the local scale
us assume that this can be performed by keeping the con-
t according to s = r 2 t. This value therefore expresses the
trast between the maximum and the minimum values con-
magnitude value that will obtained by applying the Harris-
stant. Let us therefore multiply the amplitude of the original
Laplace operator [49] to a Gaussian blob with unit con-
Gaussian blob f by a factor 2πt0 so as to obtain an input sig-
trast, provided that scale selection is performed using scale-
nal with unit contrast as measured by the range between the normalized derivatives with γ = 1. In all other respects, the
minimum and maximum values. Then, the maximum value scale selection properties of the Harris-Laplace operator are
over scale of the contrast normalized Gaussian blob will be similar to the scale selection properties of the Laplacian op-
given by erator.

 2  1
∇  = (34) 3.1.3 The Rescaled Level Curve Curvature Operator
norm L max
2
1 When applying the rescaled level curve curvature operator
(det Hnorm L)max = (35)
16 κ̃γ −norm (L) to a rotationally symmetric Gaussian blob we
1 − 4k obtain
(D1,norm L)max = (36)
16
t 2γ (x 2 + y 2 ) − 3(x 2 +y 2
1 κ̃γ −norm (L) = − e 2(t0 +t)
. (41)
(D2,norm L)max = (37) 8π(t0 + t)6
4
This expression assumes its spatial extremum on the circle
These expressions provide a way to express mutually related
magnitude thresholds for the different interest point detec- 2
x2 + y2 = (t0 + t) (42)
tors as shown in Table 1. 3
J Math Imaging Vis

where the extremum value is the scale-normalized Laplacian ∇γ2 −norm L and the scale-
normalized determinant of the Hessian det Hγ −norm L are
t 2γ given by
κ̃γ −norm (L) = − (43)
12eπ 3 (t 0 + t)5
(x 2 + y 2 − 2(t0 + t)) − 2(t
x 2 +y 2

and this entity assumes its extremum over scale at ∇γ2 −norm L = t γ e 0 +t) (48)
2π(t0 + t)3
2γ (t0 + t − x 2 − y 2 ) − xt2 +y 2

tκγ −norm (L) = t0 . (44) det Hγ −norm L = t 2γ e 0 +t (49)


5 − 2γ 4π 2 (t0 + t)5

In the special case when γ = 7/8 [4] this corresponds to With complementary Gaussian post-smoothing with scale
parameter tpost = c2 t, the resulting differential expressions
7 assume the form
tκ̃(L) = t0 ≈ 0.539 t0 ⇒ σκ̃(L) ≈ 0.734 σ0 (45)
13
∇γ2 −norm L
with the corresponding scale-normalized response +y 2 2
(x 2 + y 2 − 2(t0 + (1 + c2 )t)) − 2(t x+(1+c
 2γ =t γ
e 0
2 )t
(50)
  (5 − 2γ )5 γ t0 2π(t0 + (1 + c2 )t)3
κ̃(L) max = (46)
9375 · 41−γ e π 3 t05 5 − 2γ det Hγ −norm L
x +y 2 2
and the following approximate relation for γ = 7/8 if the (t0 + (1 + 2c2 )t − x 2 − y 2 ) − t +(1+2c
= t 2γ e 0
2 )t
(51)
Gaussian blob is normalized to unit contrast 4π 2 (t0 + t)2 (t0 + (1 + 2c2 )t)3
  0.00963 and assume their extremal scale-normalized responses over
κ̃(L) max = 1/4
(47) scale at
t0
γ t0
tˆ∇ 2 L = (52)
Due to the use of a γ -value not equal to one, this magni- (1 + c2 )(2 − γ )
tude measure is not fully scale invariant. The scale depen- 
( 1 + 2c2 + c4 (1 − γ )2 − (1 + c2 )(1 − γ ))
dency can, however, be compensated for by multiplying the tˆdet HL = t0
maximum feature response over scale by a scale-dependent (1 + 2c2 )(2 − γ )
compensation factor t 2(1−γ ) . (53)

In the specific case when γ = 1 and c = 1/2, these local


3.2 Scale Selection with Complementary Post-smoothing extrema over scale are given by

tˆ∇ 2 L = 0.800 t0 ⇒ σ̂∇ 2 L ≈ 0.894 σ0 (54)


When linking image features at different scales into fea-
ture trajectories, the use of post-smoothing of any differ- tˆdet HL ≈ 0.817 t0 ⇒ σ̂det HL ≈ 0.904 σ0 (55)
ential expression Dnorm L according to (18) was proposed
in [39] to simplify the task for the scale linking algorithm, In other words, by comparison with the results in Sect. 3.1.1,
we find that the use of a post-smoothing operation with inte-
by suppressing small local perturbations in the responses of
gration scale determined by c = 1/2, the scale estimates√ will
the differential feature detectors at any single scale. Since
be about 10 % lower when measured in units of σ = t .
this complementary post-smoothing operation will affect the
To obtain unbiased scale estimates that lead to tˆ = t0 for a
magnitude values of the scale-normalized differential re-
Gaussian blob, we can either multiply the scale estimates by
sponses that are used in the different interest point detectors,
correction factors from (52) and (53) or choose γ as function
one may ask how large effect this operation will have on the of c according to
resulting scale estimates.
In this section, we shall analyze the influence of the post- 2(1 + c2 )
γ∇ 2 L = (56)
smoothing operation for scale selection based on local ex- 2 + c2
trema over scale of scale-normalized derivatives.
2 + 3c2
γdet HL = (57)
2(1 + c2 )
3.2.1 The Laplacian and the Determinant of the Hessian
Operators With c = 1/2, the latter settings correspond to the following
values of γ :
Consider again a rotationally symmetric Gaussian blob (22) 10
γ∇ 2 L = ≈ 1.11 (58)
with its scale-space representation of the form (24). Then, 9
J Math Imaging Vis

11 3.2.3 The Hessian Feature Strength Measure II


γdet HL = = 1.10 (59)
10
For the Hessian feature strength measure II (10), we also
3.2.2 The Hessian Feature Strength Measure I have a corresponding situation with a logical switching be-
tween two differential entities |Lpp | and |Lqq | with Lpp and
To analyze the effect of the post-smoothing operation for the Lqq determined by (12) and (13). Solving for boundary be-
Hessian feature strength measure I computed for a Gaussian tween these domains, which is determined by Lpp + Lqq =
blob, which is given by 0, gives that we should select |Lqq | within the circular re-
gion
D1,γ −norm L
   x 2 + y 2 ≤ 2(t0 + t) (64)
= t 2γ (t0 + t) t + t0 − x 2 − y 2
2 2
and |Lpp | outside. Solving for the corresponding integrals
− xt +y gives
 2  e 0+t
+ κ 2(t0 + t) − x 2 − y 2 (60)
4π 2 (t0 + t)6 Lqq inside
   
provided that this entity is positive, let us initially disregard = −t γ (t0 + t) t0 + 1 − c2 t
the effect of the local condition det HL − k trace2 HL > 0
    −1− t0 +t
in (8) and integrate the closed-form expression (60) over the + (t0 + t) t0 + 1 + 3c2 t e c2 t

entire image plane R2 instead of over only the finite region     2 


where this entity is positive / 2π(t0 + t)2 t0 + 1 + c2 t (65)
√ −
t+t0
−1
1 − 4κ − (1 − 4κ) t γ e c2 t
x +y ≤
2 2
(t0 + t) (61) Lpp outside = − (66)
2κ 2π(t0 + t)(t0 + (1 + c2 )t)

Then, complementary post-smoothing with integration scale with


tpost = c2 t implies that this approximation of the post-
D2 L = |Lqq inside | + |Lpp outside | (67)
smoothed differential entity is given by
Unfortunately, it is again hard to solve for the local ex-
(D1,γ −norm L)(0, 0; t) trema over scale of the post-smoothed derivative expres-
  sions in closed form. For this reason, let us approximate the
≈ t 2γ 1 − 4κ − 2(1 − 4κ) c2 − 8κ c4 t 2
   composed expression D2 L by the contribution from its first
+ 2 1 + c2 (1 − 4κ)t t0 + (1 − 4κ)t02 term4 Lqq inside and with the integral extended from the cir-
    3  cular region (64) to the entire image plane
/ 4π 2 (t0 + t)3 t0 + 1 + 2c2 t (62)
t γ (t0 + (1 − c2 )t)
Lqq = − (68)
Corresponding integration within the finite support region 2π(t0 + t)(t0 + (1 + c2 )t)
(61) where D1 > 0 gives an expression that is too complex
to be written out here. Then, the local extrema over scale are given by the solutions
Unfortunately, it is hard to analyze the scales at which of the third-order equation
these entities assumes local extrema over scale, since differ-  4   
c − 1 (γ − 2)t 3 + tt02 3c2 − 3γ + 2
entiation of the above mentioned expression and solving for  
its roots leads to fourth-order equations. In the case of γ = 1, + t 2 t0 c4 (γ − 1) + 3c2 − 3γ + 4 − γ t03 = 0 (69)
c = 1/2 and κ = 0.04, we can, however, find the numerical
where the special case with γ = 1 and c = 1/2 has the nu-
solution
merical solution
tˆ ≈ 0.813 t0 ⇒ σ̂ ≈ 0.902 σ0 (63) tˆ ≈ 0.699 t0 ⇒ σ̂ ≈ 0.836 σ0 (70)
For these parameter settings, the use of a spatial post- For the D2,norm L operator and these parameter values, the
smoothing operation does again lead to scale estimates that use of a spatial post-smoothing operation does therefore lead
are about 10 % lower.
If we restrict ourselves to the analysis of a single isolated 4 This approximation may be reasonable for small values of c for which
Gaussian blob, a similar approximation holds for the signed the major contribution of the post-smoothing integration originates
Hessian feature strength measure D̃1,γ −norm L. from values of D2 L near the interest point.
J Math Imaging Vis

to scale estimates that are about 16 % lower, and the influ- 3.3 Influence of Affine Image Deformations
ence is therefore stronger than for the Laplacian ∇norm
2 L, de-
terminant of the Hessian det Hnorm L or the Hessian feature To analyze the behaviour of the different interest point de-
strength D1,norm L operators. tectors under image deformations, let us next consider an
If we restrict ourselves to the analysis of a single isolated anisotropic Gaussian blob as a prototype model of a rota-
Gaussian blob, a similar approximation holds for the signed tionally symmetric Gaussian blob that has been subjected to
Hessian feature strength measure D̃2,γ −norm L. an affine image deformation that we can see as representing
a local linearization of the perspective mapping from a sur-
3.2.4 The Rescaled Level Curve Curvature Operator face patch in the world to the image plane. Specifically, we
can model the effect of foreshortening by different spatial
extents t1 and t2 along the different coordinate directions
If we apply post-smoothing to the rescaled level curve cur-
vature computed for a rotationally symmetric Gaussian blob 1 2
− 2t
x y2
− 2t
(41) with post-smoothing scale tpost = c2 t, we obtain f (x, y) = g(x; t1 ) g(y; t2 ) = √ e 1 2 (76)
2π t1 t2

κ̃γ −norm (L) where the ratio between the scale parameter t1 and t2 is re-
lated to the angle θ between the normal directions of the
t 2γ ((t0 + t)(x 2 + y 2 ) + 2c2 t (t0 + t) + 6c4 t 2 ) surface patch and the image plane according to
=−
8π 3 (t0 + t)4 (t0 + (1 + 3c2 )t)3 
σ2 t2
− 3(x 2 +y 2 ) = = cos θ (77)
×e 2(t0 +(1+3c2 )t)
(71) σ1 t1
if we without loss off generality assume that t1 ≥ t2 . Since
This entity assumes it spatial extremum on the circle
all the feature detectors we consider are based on rotation-
2 (t0 + t)2 − 9c4 t 2 ally invariant differential expressions, it is sufficient to study
x2 + y2 = (72) the case when the anisotropic Gaussian blob is aligned to
3 (t0 + t)
one the coordinate directions. Due to the semi-group prop-
and the extremum value on this circle is erty of the (one-dimensional) Gaussian kernel, the scale-
space representation of f is then given by
κ̃γ −norm (L)extr
L(x, y; t) = g(x; t1 + t) g(y; t2 + t)
t 2γ
=− 1 2
− x − y
2
12π 3 (t0 + t)3 (t0 + (1 + 3c2 )t)2 = √ e 2(t1 +t) 2(t2 +t) (78)
2 2
2π (t1 + t)(t2 + t)
− 3(x +y 2)
×e 2(t0 +(1+3c )t)
(73) Note on Relation to Influence Under General Affine Trans-
formations A general argument for studying the influence
By differentiating this expression with respect to the scale
of non-uniform scaling transformations can be obtained by
parameter t, it follows that the selected scale level will be a decomposing a general two-dimensional affine transforma-
solution of the third-order equation tion matrix A into [32]
   
1 + 3c2 (5 − 2γ )t 3 + 10 − 9c4 − 6γ + 6c2 (3 − 2γ ) t0 t 2 A = R1 diag(σ1 , σ2 ) R2−1 (79)
 
+ 5 + c2 (3 − 6γ ) − 6γ t0t t − 2γ t03 = 0. (74) where R1 and R2 can be forced to be rotation matrices, if we
relax the requirement of non-negative entries in the diagonal
Unfortunately, the closed form expression for the solution
elements σ1 and σ2 of a regular singular value decomposi-
is rather complex. Nevertheless, we can note that due to the
tion. With this model, the geometric average of the absolute
homogeneity of this equation, the solution will always be values of the diagonal entries
proportional the scale t0 of the original Gaussian blob. In 
the specific case with γ = 7/8 and c = 1/2 we obtain σuniform = |σ1 σ2 | (80)

tκ̃γ −norm (L) ≈ 0.493 t0 ⇒ σκ̃γ −norm (L) ≈ 0.702 σ0 (75) corresponds to a uniform scaling transformation. We know
that the Gaussian scale-space is closed under uniform scal-
In other words, compared to the case without post-smooth- ing transformations, rotations and reflections.
ing (45), the relative difference between the selected scale The differential expressions we use for detecting inter-
levels√is here less than 5 %, when measured in units of est points are based on rotationally invariant differential in-
σ = t. variants, which implies that the scale estimates will also be
J Math Imaging Vis

rotationally invariant. Furthermore, our scale estimates are For a general value of γ , the explicit solution is too complex
transformed in a scale covariant way under uniform scal- to be written out here. In the specific case of γ = 1, however,
ing transformations. Hence, if we without essential loss of we obtain for t1 ≥ t2
generality disregard reflections and assume that σ1 and σ2 
are both positive, the degree of freedom that remains to be 1
t∇ 2 L = 222/3 R2 − 4(t1 + t2 )
studied concerns non-uniform scaling transformations of the 24
form √ 
8 3 2(t12 − 10t1 t2 + t22 )
   − (86)
1 σ1 σ2 R4
√ diag(σ1 , σ2 ) = diag ,
σ1 σ2 σ2 σ1
  where
√ 1

= diag s, √ (81)
s R1 = 9t14 + 1518t12 t22 + 9t24 (87)
whose influence on the scale estimates will be investigated

in this section. R2 = 3 7t13 + 57t12 t2 + 3(t1 − t2 ) + R1 + 57t1 t22 + 7t23 (88)


3.3.1 The Laplacian operator R3 = 9t14 + 1518t12 t22 + 9t24 (89)


For the Laplacian operator, the γ -normalized response as R4 = 3 7t13 + 57t12 t2 + 3(t1 − t2 )R3 + 57t1 t22 + 7t23 (90)
function of space and scale is given by
 2  which in the special case of t1 = t2 = t0 reduces to
∇γ −norm L (x, y; t)
t∇ 2 L = t 0 (91)
t γ e−x /2(t1 +t)−y /2(t2 +t)
2 2

=−
2π((t1 + t)(t2 + t))5/2 If we on the other hand reparameterize the scale parame-
   ters t1 and t2 of the Gaussian blob as t1 = s t0 and t2 = t0 /s,
× 2t 3 + t 2 3t1 + 3t2 − x 2 − y 2
corresponding to a non-uniform scaling transformation with
 
+ t t12 + 4t1 t2 − 2t1 y 2 + t22 − 2t2 x 2 relative scaling factor s > 1 renormalized such the determi-
 nant of the transformation matrix is equal to one, then a Tay-
+ t1 t2 (t1 + t2 ) − t22 x 2 − t12 y 2 (82) lor expansion of t∇ 2 L around s = 1 gives
This entity has critical points at the origin (x, y) = (0, 0) 
1 1
and at t∇ 2 L = 1 − (s − 1)2 + (s − 1)3
4 4
   
(t1 + t)(t1 + 3t2 + 4t) 1  
(x, y) = ± ,0 (83) − (s − 1) + O (s − 1)
4 5
t0 (92)
(t2 + t) 4
  
(t2 + t)(3t1 + t2 + 4t) From this result we get an approximate expression for how
(x, y) = 0, ± (84) the Laplacian scale selection method is affected by affine
(t1 + t)
transformations outside the similarity group. Specifically,
where the first pair of roots corresponds to saddle points if we can note that the scales selected from local extrema over
t1 > t2 , while the other pair of roots correspond to local ex- scale of the scale-normalized Laplacian operator are not in-
trema. Unfortunately, the critical points outside the origin variant under general affine transformations.
lead to rather complex expressions. We shall therefore fo-
cus on the critical point at the origin, for which the selected 3.3.2 The Determinant of the Hessian
scale(s) will be the root(s) of the third-order equation
  By differentiation of (78) it follows that the scale-normalized
∂t ∇γ2 −norm L(0, 0; t)
determinant of the Hessian is given by
t γ −1
=
4π((t + t1 )(t + t2 ))5/2 (det Hγ −norm L)(x, y; t)
    
× −4(γ − 2)t 3 − 2(3γ − 4)t 2 (t1 + t2 ) = t 2γ t 2 + t t1 + t2 − x 2 − y 2 + t1 t2 − t2 x 2 − t1 y 2
 
+ t (3 − 2γ )t12 + (2 − 8γ )t1 t2 + (3 − 2γ )t22 − x2
− y2

 e t+t1 t+t2
× (93)
− 2γ t1 t2 (t1 + t2 ) = 0 (85) 4π (t1 + t)3 (t2 + t)3
2
J Math Imaging Vis

This expression does also have multiple critical points. = t 2γ (1 − 4k)t 2 + (1 − 4k)t (t1 + t2 )
Again, however, we focus on the central point (x, y) = (0, 0), 
for which the derivative with respect to scale is of the form − k(t1 + t2 )2 + t1 t2
 
   / 4π 2 (t1 + t)3 (t2 + t)3 (99)
∂t det Hγ2 −norm L (0, 0; t)

t 2γ −1 ((γ − 2)t 2 + (γ − 1)t (t1 + t2 ) + γ t1 t2 ) provided that this entity is positive. If we differentiate this
= (94) expression with respect to the scale parameter and set the
2π 2 (t1 + t)3 (t2 + t)3
derivative to zero, we obtain a fourth-order equation, which
This equation has a positive root at in principle can be solved in closed form, but leads to very
complex expressions, even when restricted to γ = 1.
tdet HL If we reparameterize the scale parameters according to
1 
t1 = s t0 and t2 = t0 /s and then restrict the parameter k
= 4(2 − γ )γ t1 t2 + (γ − 1)2 (t1 + t2 )2 in D1,γ −norm to k = 0.04, however, we can obtain a man-
2(2 − γ )
 ageable expression for the Taylor expansion of the selected
+ (γ − 1)(t1 + t2 ) (95) scale tD1 L as function of the non-uniform scaling factor s in
which in the special case of γ = 1 simplifies to the affine the specific case of γ = 1
covariant expression 
1 1
√ tD 1 L = 1 + (s − 1)2 − (s − 1)3
tdet HL = t1 t2 (96) 21 21

109  
Notably, if we again reparameterize the scale parameters ac- + (s − 1) + O (s − 1)
4 5
t0 (100)
1764
cording to t1 = s t0 and t2 = t0 /s, then for any non-uniform
scaling transformation renormalized such that the determi- From this expression we can see that the scales selected
nant of the transformation matrix is one, it holds that from local extrema over scale of the scale-normalized Hes-
sian feature strength measure I are not invariant under non-
tdet HL = t0 (97) uniform scaling transformations. For values of s reasonably
which implies that (in this specific case and with γ = 1) close to one, however, the deviation from affine invariance
scale selection based on the scale-normalized determinant of is quite low, and significantly smaller than for the Lapla-
the Hessian leads to affine covariant scale estimates for the cian operator (92). This could also be expected, since a
Gaussian blob model. In this respect, there is a significant major contribution to the Hessian feature strength measure
difference to scale selection based on the scale-normalized D1,γ −norm originates from the affine covariant determinant
Laplacian, for which the scale estimates will be biased ac- of the Hessian det Hγ −norm L.
cording to (86) and (92).
3.3.4 The Hessian Feature Strength Measure II
For other values of γ , a Taylor expansion of tdet HL
around s = 1 gives
With t1 > t2 , the Hessian feature strength measure II at the

1 1 origin is given by
tdet HL = 1 − (1 − γ )(s − 1)2 + (1 − γ )(s − 1)3
2 2
(D2,γ −norm L)(0, 0; t) = Lξ ξ (0, 0; t)
1 
− 2 + γ − 4γ 2 + γ 3 (s − 1)4 tγ
8 =− √ .
 2π(t1 + t) (t1 + t)(t2 + t)
  γ
+ O (s − 1)5
t0 (98) (101)
2−γ
This entity assumes its local extremum over scale at
implying a certain dependency on the relative scaling fac-
tor s. Provided that |γ − 1| < 1/2, this dependency will, 1 
however, be lower than for the Laplacian scale selection tD 2 L = 2γ (t1 + t2 ) − t1 − 3t2
4(2 − γ )
method (92) with γ = 1.

 2 
+ 16(2 − γ )γ t1 t2 + (2γ − 1)t1 − (3 − 2γ )t2
3.3.3 The Hessian Feature Strength Measure I
(102)
For the Hessian feature strength measure I, the behaviour of which in the case when γ = 1 reduces to
the scale-normalized response at the origin is given by

1 2
D1,γ −norm (0, 0; t) tD 2 L = t1 + 14t1 t2 + t22 + t1 − t2 (103)
4
J Math Imaging Vis

If we again reparameterize the scale parameters according to (5 − 4γ )(25 + 30γ − 8γ 2 )


− (s − 1)4
t1 = s t0 and t2 = t0 /s and in order to obtain more compact 500
expressions restrict ourselves to the case when γ = 1, then 
  2γ
a Taylor expansion of tD2 L around s = 1 gives + O (s − 1) 5
t0 (110)
5 − 2γ

1 1 1 which in the case with γ = 7/8 assumes the form
tD2 L = 1 + (s − 1) − (s − 1)2 + (s − 1)3
2 8 8 
 3 3
13   tκ(L) = 1 − (s − 1)2 + (s − 1)3
− (s − 1)4 + O (s − 1)5 t0 (104) 20 20
128 
1083   7
From a comparison with (92), (97) and (100) we can see that − (s − 1)4 + O (s − 1)5 t0 (111)
8000 13
the scales selected from the scale-normalized Hessian fea-
ture strength measure II are more sensitive to non-uniform with second- and third-order relative bias terms about
scaling transformations than the scales selected by the scale- three times the magnitude compared to the Hessian feature
normalized Laplacian ∇norm
2 L, the determinant of the Hes- strength measure I in (100).
sian det Hnorm L or the Hessian feature strength measure
D1,norm .
4 Scale Selection by Weighted Averaging Along Feature
3.3.5 The Rescaled Level Curve Curvature Operator
Trajectories
For the anisotropic Gaussian blob model, a computation of
the rescaled level curve curvature operator κγ −norm (L) gives The treatment so far has been concerned with scale selec-
tion based on local extrema over scale of scale-normalized
κγ −norm (L) derivatives. Concerning the new scale selection method
based on weighted averaging over scale of scale-normalized
t 2γ (t (x 2 + y 2 ) + t1 y 2 + t2 x 2 ) − 32 ( t x+t
2 2
+ t y+t )
=− e 1 2 (105) derivative responses, an important question concerns how
8π 3 ((t1 + t)(t2 + t))7/2 the scale estimates from this new scale selection method are
This entity assumes its spatial maximum on the ellipse related to the scale estimates from the previously established
scale selection based on local extrema over scale. In this sec-
x2 y2 2 tion, we will present a corresponding theoretical analysis of
+ = (106)
t1 + t t2 + t 3 weighted scale selection concerning the basic questions:
and on this ellipse it holds that – How are the scale estimates related between the different
interest point detectors? (Sect. 4.1)
t 2γ
κγ −norm (L) = − (107) – How much does the post-smoothing operation influence
12eπ 3 ((t1 + t)(t2 + t))5/2
the scale estimates? (Sect. 4.2 )
By differentiating this expression with respect to t and set- – How are the scale estimates influenced by affine image
ting the derivative to zero, it follows that the extremum over deformations? (Sect. 4.3)
scale is assumed at
Given that image features x(t) at different scales t have been
1  linked into a feature trajectory T over scale
tκ(L) = −(5 − 4γ )(t1 + t2 )
4(5 − 2γ )   

 T = x(t); t : t ∈ [tmin , tmax ] (112)
+ 32γ (5 − 2γ )t1 t2 + (5 − 4γ )2 (t1 + t2 )2 (108)
scale selection by weighted averaging over scale implies that
which in the special case when γ = 5/4 reduces to the affine the scale estimate is computed as [39]
covariant expression 
√ τ ψ(|(Dγ −norm L)(x(τ ); τ )|) dτ
tκ(L) = t1 t2 (109) τ̂T = τ ∈T
τ ∈T ψ(|(Dγ −norm L)(x(τ ); τ )|) dτ
By again reparameterizing the scale parameters in the Gaus-  tmax dt
t=tmin (log t) ψ(|(Dγ −norm L)(x(t); t)|) t
sian blob model according to t1 = s t0 and t2 = t0 /s and per- =  tmax (113)
dt
forming a Taylor expansion around s = 1 for a general value t=tmin ψ(|(Dγ −norm L)(x(t); t)|) t
of γ , it follows that
 for some positive and monotonically increasing trans-
(5 − 4γ ) (5 − 4γ ) formation function ψ of the magnitude values
tκ(L) = 1 − (s − 1)2 + (s − 1)3
10 10 |(Dγ −norm L)(x(τ ); τ )| of the differential feature responses.
J Math Imaging Vis

Specifically, the following family of scale invariant transfor- between the Hessian feature strength measure II (28) and
mation functions was considered the Laplacian response (27) at the center of a Gaussian blob,
  the scale estimates for D1,γ −norm L and D2,γ −norm L will be
ψ |Dγ −norm L| = wDL |Dnorm L|a (114)
analogous:
where wDL ∈ [0, 1] is a so-called feature weighting func-
∞ (1−4k) t 2 dt (1−4k) log t0
tion that measures the relative strength of the feature detec- t=0 4π 2 (t0 +t)4 log t t 24π 2 t0
tor DL compared to other possible competing types of fea- τ̂D1 L =  ∞ (1−4k) t 2 dt = (1−4k)
= log t0
ture detectors and a is the scalar parameter in the self-similar t=0 4π 2 (t0 +t)4 t 24π 2 t02
power law. (118)
In this section, we shall analyze the scale selection prop- ∞ t dt log t0
erties of this construction for the differential feature detec- t=0 2π(t0 +t)2 log t t 2πt0
τ̂D2 L = ∞ t dt
= 1
= log t0 (119)
tors defined in Sect. 2.2 under the simplifying assumptions t=0 2π(t0 +t)2 t 2πt0
of wDL = 1 and a = 1. With respect to the analysis at the
center of a Gaussian blob, the assumption of wDL = 1 is
When expressed in terms of the regular scale parameter
particularly relevant for the weighting functions of the form

L2ξ ξ + 2L2ξ η + L2ηη tˆ = eτ̂ = t0 (120)


wD L = (115)
A(L2ξ + L2η ) + L2ξ ξ + 2L2ξ η + L2ηη + ε2
the weighted scale selection method does hence for a rota-
considered in [39] if we make use of the fact that Lξ = Lη = tionally symmetric Gaussian blob lead to similar scale esti-
0 at any critical point (as at the center of a Gaussian blob) mates as are obtained from local extrema over scale of γ -
and disregard the influence of the noise suppression param-
normalized derivatives (29) when γ = 1.
eter ε.
Since these scale estimates are similar to the scale esti-
4.1 The Pure Second-Order Interest Point Detectors mates obtained form local extrema over scale, it follows that
the scale-normalized magnitude values will also be similar
From the explicit expression for the magnitude of the scale- and the relationships between scale-normalized thresholds
normalized Laplacian response at the center of rotationally described in Table 1 will also hold for scale selection based
symmetric Gaussian blob (25) it follows that the weighted on weighted averaging over scale.
scale selection estimate according to (113) will in the case
of γ = 15 and with the effective scale parameter τ defined
as τ = log t be given by Corresponding Scale Estimates for General Values of γ
∞ For a general value of γ ∈]0, 2[, the corresponding scale
t dt log t0
t=0 π(t +t)2 log t t πt0
estimates become as follows in terms of effective scale
τ̂∇ 2 L =  ∞ 0 t dt
= 1
= log t0 (116) τ = log t:
t=0 π(t0 +t)2 t πt0

Similarly, from the explicit expression for the determinant π(1 − γ ) cot(πγ ) + 1
τ̂∇ 2 L = τ̂D2 L = log t0 − (121)
of the Hessian at the center of the Gaussian blob (26), it 1−γ
follows that the weighted scale estimate will be determined τ̂det HL
by
∞ = τ̂D1 L
t2 dt log t0
t=0 4π 2 (t0 +t)4 log t t 24π 2 t0 π csc(2πγ )   
τ̂det HL = ∞ = 1
= log t0 (117) = −2 4(γ − 3)γ + 11 γ log t0
t2 dt
t=0 4π 2 (t0 +t)4 t 24π 2 t02 Γ (4 − 2γ ) Γ (2γ )
+ 2π(γ − 1)(2γ − 3)(2γ − 1) cot(2πγ )
Due to the similarity between the explicit expressions for

the Hessian feature strength measure I (27) and the determi- + 6 log t0 − 12(γ − 2)γ − 11 (122)
nant of the Hessian response (26) as well as the similarity
where both expressions have the limit value log t0 when
5 In this section we will in many cases restrict the analysis to the spe- γ → 1. (Note that | cot(πγ )| → ∞ when γ → 1.)
cific case of γ = 1, since some of the results become significantly more By comparing these scale estimates to the corresponding
complex for a general value of γ = 1. In a few cases where the corre- γ
sponding results become reasonably compact, we will, however, in- scale estimate τ̂ = log 2−γ in (29) obtained from local ex-
clude them. trema over scale, we can compute a Taylor expansions of the
J Math Imaging Vis

difference in the scale estimates: for γ = 1 given by


   ∞ 2t (t0 +(1+c2 )t) dt
γ t=0 2π(t0 +(1+c2 )t)3 log t t
τ̂∇ 2 L − log τ̂∇ 2 L =  ∞ 2t (t0 +(1+c2 )t) dt
2−γ
t=0 2π(t0 +(1+c2 )t)3 t
 2 
π 1 4 
= − 2 (γ − 1) + π − 30 (γ − 1)3
t0
log( )
1+c2
3 45 πt0 (1+c2 )
  = 1
+ O (γ − 1)4 πt0 (1+c2 )
   
≈ 1.290 (γ − 1) + 1.498 (γ − 1)3 + O (γ − 1)4 (123) t0
  = log (125)
γ 1 + c2
τ̂det HL − log ∞
2−γ t 2 (t0 +(1+2c2 )t)
t=0 log t 4π 2 (t0 +t)2 (t0 +(1+2c2 )t)3
 2   4  τ̂det HL = ∞
2π 8π 50 t 2 (t0 +(1+2c2 )t)
= − 6 (γ − 1) + − (γ − 1)3 t=0 4π 2 (t0 +t)2 (t0 +(1+2c2 )t)3
3 45 3
  (2c2 −(1+c2 ) log(1+2c2 )) log 1+2c
2
+ O (γ − 1)4 2t0

  =
32c6 π 2 t02
≈ 0.580 (γ − 1) + 0.651 (γ − 1)3 + O (γ − 1)4 (124) (2c −(1+c2 ) log(1+2c2 ))
2

16c6 π 2 t02
 
Notably, the difference in scale estimates between the two t0
= log √ (126)
types of scale selection approaches is smaller6 for scale 1 + 2c2
selection using the determinant of the Hessian det HL or
which agree with the corresponding scale estimates (52)
the Hessian feature strength measure D1,norm L compared to and (53) for local extrema over scale of post-smoothed γ -
scale selection based on the Laplacian ∇norm
2 L or the Hes- normalized derivative expressions.
sian feature strength measure D2,norm L. In the specific case when c = 1/2 these scale estimates
reduce to

4.2 Influence of the Post-smoothing Operation tˆ∇ 2 L = 0.8 t0 ⇒ σ̂∇ 2 L ≈ 0.894 σ0 (127)
tˆdet HL ≈ 0.816 t0 ⇒ σ̂det HL ≈ 0.904 σ0 (128)
4.2.1 The Laplacian and the Determinant of the Hessian again agreeing with our previous results (54) and (55) re-
Operators garding scale selection based on local extrema over scale of
post-smoothed scale-normalized derivative expressions.

From the explicit expressions for the post-smoothed Lapla- 4.2.2 The Hessian Feature Strength Measure I
cian (50) and the post-smoothed determinant of the Hes-
sian (51), it follows that the weighted scale estimates are To analyze the effect of weighted scale selection for the Hes-
sian feature strength measure I, corresponding weighted in-
tegration over scale of the approximation (62) of the post-
smoothed differential entity D1 L for γ = 1 gives
6A plausible explanation why the difference between the scale esti- ∞
mated is smaller for the determinant of the Hessian det HL and the (D1,γ −norm L) log t dtt
Hessian feature strength measure D1,norm L compared to difference τ̂D1 L ≈ t=0
∞ dt
= θD1 L + log t0 (129)
in scale estimates for the Laplacian ∇norm
2 L and the Hessian feature t=0 (D1,γ −norm L) t
strength measure D2,norm L is that second-order derivative responses
where
are squared for the determinant of the Hessian det HL and the Hessian
feature strength measure D1,norm L, whereas the Laplacian ∇norm
2 L and ND1 L
the Hessian feature strength measure D2,norm L operators depend on θD1 L = − (130)
DD 1 L
the second-order derivative responses in a linear way.
Thereby, the integrals that define the weighted scale selection es- and
timates will get a comparably higher relative contribution from scale    
levels near the maximum over scale, which in turn implies that the ND1 L = 1 + 3c2 + 2c4 (1 + 2κ) log2 1 + 2c2
influence due to skewness in the scale-space signature caused by val-    
ues of γ = 1 will be lower (compare with Sect. 5.1.1). By varying the − 2c2 1 + 2κ + 2c2 (1 + κ) log 1 + 2c2
power a in the self-similar transformation function (114), it is more
generally possible to modulate this effect. − 8c6 κ (131)
J Math Imaging Vis
   
DD1 L = 2 1 + 3c2 + 2c4 (1 + 2κ) log 1 + 2c2 and the scale estimate obtained by weighted scale selection
  is given by
− 2c2 1 + 2κ + 2c2 (1 + 2κ) + 2c4 κ (132)
∞ t (2t+t1 +t2 ) dt
In the specific case of c = 1/2 and κ = 0.04, this scale esti- t=0 2π (t1 +t)3/2 (t2 +t)3/2 log t t
τ̂∇ 2 L = ∞ t (2t+t1 +t2 ) dt
mate is given by
t=0 2π (t1 +t)3/2 (t2 +t)3/2 t

tˆD1 L ≈ 0.813 t0 ⇒ σ̂D1 L ≈ 0.902 σ0 (133) log( t


4t1 t2

1 +t
√2
+2 t1 t2 )  
π t1 t2 4t1 t2
in agreement with our previous result (63) regarding scale = = log √ (138)
√1 t 1 + t2 + 2 t1 t 2
selection based on local extrema over scale of post-smoothed π t1 t2

scale-normalized derivative expressions.


With a reparameterization of the scale parameters t1 and t2
of the Gaussian blob as t1 = s t0 and t2 = t0 /s, correspond-
4.2.3 The Hessian Feature Strength Measure II
ing to a non-uniform scaling transformation with relative
scaling factor s > 1 renormalized such the determinant of
To analyze the effect of post-smoothing of the Hessian fea-
the transformation matrix is equal to one, the scale estimate
ture strength measure II in (10), let us again approximate
the composed post-smoothed differential expression in (67) in units of t can be written
by the contribution from its first term (68) with the spatial 4t1 t2 1 √
integration extended to the entire plane. Then, the weighted tˆ∇ 2 L = eτ̂∇ 2 L = √ = t1 t2
t1 + t2 + 2 t1 t2 1 +t2
t√
1
2 + 4 t1 t2
scale estimate can be approximated by
∞ 4s
Lqq log t dtt = t0 (139)
τ̂D2 L ≈ t=0
∞ = θD2 L + log t0 (134) (1 + s)2
dt
t=0 Lqq t
Notably, this scale estimate is not identical to the scale esti-
where mate (86) obtained from local extrema over scale. The Tay-
lor expansion of t∇ 2 L around s = 1 is in turn given by
log(1 + c2 )((1 + c2 ) log(1 + c2 ) − 4c2 )
θD2 L = − (135) 
2((1 + c2 ) log(1 + c2 ) − 2c2 ) 1 1
t∇ 2 L = 1 − (s − 1)2 + (s − 1)3
In the specific case when c = 1/2 this scale estimate reduces 4 4

to 3  
− (s − 1) + O (s − 1)
4 5
t0 (140)
16
tˆD2 L ≈ 0.694 t0 ⇒ σ̂D1 L ≈ 0.834 σ0 (136)
and is, however, similar until the third-order terms to the
in agreement with our previous result (70) regarding scale
Taylor expansion (92) of the corresponding scale estimate
selection based on local extrema over scale of post-smoothed
obtained from local extrema over scale. In this respect, the
scale-normalized derivative expressions.
behaviour of the two scale selection methods is qualitatively
rather similar when applied to the anisotropic Gaussian blob
4.3 Influence of Affine Image Deformations
model.
To analyze how the scale estimates tˆ obtained by weighted
averaging along feature trajectories are affected by affine 4.3.2 The Determinant of the Hessian
image deformations, let us again consider an anisotropic
Gaussian blob (76) as a prototype model of a rotationally At the origin, the response of the determinant of the Hessian
symmetric Gaussian blob that has been subjected to an affine operator (93) simplifies to
image deformation and with its scale-space representation
according to (78). t2
(det Hnorm L)(0, 0; t) = (141)
4π 2 (t1 + t)2 (t2 + t)2
4.3.1 The Laplacian Operator
and the scale estimate obtained by weighted scale selection
At the origin, the scale-normalized Laplacian response ac- is
cording to (82) reduces to ∞ t2 dt
t=0 4π 2 (t1 +t)2 (t2 +t)2 log t t
 2  t (2t + t1 + t2 ) τ̂det HL = ∞
∇norm L (0, 0; t) = − (137) t2 dt
2π (t1 + t)3/2 (t2 + t)3/2 t=0 4π 2 (t1 +t)2 (t2 +t)2 t
J Math Imaging Vis

(t1 +t2 )(log2 t1 −log2 t2 )−2(t1 −t2 ) log(t1 t2 ) which simplifies to the following form for κ = 0.04
8π 2 (t1 −t2 )3
= (t1 +t2 )(log t1 −log t2 )−2(t1 −t2 ) 
1 1
4π 2 (t1 −t2 )3 tD1 L = 1 + (s − 1)2 − (s − 1)3
21 21
log(t1 t2 ) 
= (142) 559  
2 + (s − 1) + O (s − 1)
4 5
t0 (149)
8820
corresponding to the affine covariant scale estimate
√ and agreeing until the third-order terms with the correspond-
tˆdet HL = eτ̂det HL = t1 t2 (143) ing Taylor expansion (100) for the scale estimate obtained
from local extrema over scale.
and in agreement with our earlier result (96) for scale selec- Specifically, a comparison with the corresponding ex-
tion from local extrema over scale. pression for the Laplacian operator (140) shows that scale
selection based on the Hessian feature strength measure I
4.3.3 The Hessian Feature Strength Measure I is less sensitive to affine image deformations compared to
scale selection based on the Laplacian.
With the Hessian feature strength measure I at the origin
given by (99), the scale estimate obtained by weighted scale 4.3.4 The Hessian Feature Strength Measure II
selection is determined by
Assuming that t1 ≥ t2 , the Hessian feature strength measure
∞
(D1,γ −norm L) log t dtt ND 1 L II at the origin is given by
τ̂D1 L = t=0∞ dt
= (144)
t=0 (D1,γ −norm L) t
DD 1 L
t
(D2,norm L)(0, 0; t) = Lxx (0, 0; t) = − √ 3√
where 2π t1 + t t2 + t
(150)
ND1 L = κ(t1 − t2 )3
  and the weighted scale estimate
− (1 − κ)t1 t2 (t1 + t2 ) log2 t1 − log2 t2
  ∞ t dt
− t2 (t2 − t1 ) log(t1 ) (2 − 3κ)t1 + κt2 t=0 2π √t +t 3 √t +t log t t
  τ̂D2 L =  ∞ 1
t
2
dt
+ t1 (t1 − t2 ) log(t2 ) κ(t1 − 3t2 ) + 2t2 t=0 2π √t +t 3 √t +t t
1 2
    √ √ √ √
DD1 L = (t1 − t2 ) κ t12 − 6t1 t2 + t22 + 4t1 t2 2 t2 log(t1 t2 )+( t1 + t2 )(log 16−2 log(t1 +t2 +2 t1 t2 ))

2π t1 (t1 −t2 )
− 2(1 − κ)t1 t2 (t1 + t2 )(log t1 − log t2 ) =− 1√
π(t1 + t1 t2 )
With a reparameterization of the scale parameters according (151)
to t1 = s t0 and t2 = t0 /s, this expression simplifies to
With the scale parameters reparameterized according to t1 =
N s t0 and t2 = t0 /s, the corresponding scale estimate can be
D1 L
τ̂D1 L = (145) written
DD 1L √ √
 √ √  √t1 +√t2
( t1 + t2 )2 t1 − t2 √
with tˆD2 L = eτ̂D2 L = √ t1 t2
4 t1 t2
 2   

ND = κ 1 − s2 1 + s 2 log s + 1 − s 2 (146)   1+s
1L 4s 1−s


  = t0 (152)
DD = 4(1 − κ) 1 + s 2 s 2 log s (1 + s)2
1L
     for which a Taylor expansion around s = 1 gives
+ 1 − s 2 κ 1 − 6s 2 + s 4 + 4s 2 (147)

1 1 1
A Taylor expansion around s = 1 of the scale estimate ex- tD2 L = 1 + (s − 1) − (s − 1)2 + (s − 1)3
2 8 12
pressed in units of t = exp t gives 
 23  
κ κ − (s − 1)4 + O (s − 1)5 t0 (153)
tD 1 L = 1 + (s − 1)2 − (s − 1)3 384
1 − 4κ 1 − 4κ
 and agreeing until the second-order terms with the corre-
25κ − 66κ 2  
+ (s − 1)4
+ O (s − 1)5
t0 (148) sponding Taylor expansion (104) for the scale estimate ob-
20 (1 − 4κ)2 tained from local extrema over scale.
J Math Imaging Vis

Again, the scale estimates for scale selection based on at the center of the blob (below assuming no post-smoothing
the Hessian feature strength measure II are more affected by corresponding to c = 0):
affine image deformations compared to the scale estimates
 2  t
obtained by the determinant of the Hessian, the Hessian fea- ∇norm L (0, 0; t) = − (154)
ture strength measure I or the Laplacian. π(t0 + t)2
t2
(det Hnorm L)(0, 0; t) = (155)
4π 2 (t0 + t)4
5 Relations Between the Scale Selection Methods
The left column in the upper and middle rows in Fig. 3
5.1 Rotationally Symmetric Gaussian Blob shows these graphs with a linear scaling of the regular
scale parameter t and the right column shows correspond-
From the above mentioned results, we can first note that for ing graphs with a logarithmic scaling of the scale parameter
the specific case of a rotationally symmetric Gaussian blob, in terms of effective scale τ .
the scale estimates obtained from local extrema over scale As can be seen from the latter graphs, the scale-space sig-
vs. weighted averaging over scale are very similar. natures assume a symmetric shape when expressed in terms
Table 2 shows the scales that are selected for the Lapla- of effective scale, which implies that the weighted scale es-
cian ∇norm
2 L and the determinant of the Hessian det Hnorm L timates, which correspond to the center of gravity of the
in the presence of a general post-smoothing operation. Ta- graphs, will be assumed at a similar position as the global
ble 3 shows corresponding approximate estimates for the extremum over scale. This property can also be understood
Hessian feature strength measure D1,norm L and the Hessian algebraically, due to the functional symmetry of (154) and
feature strength measure D2,norm L for c = 1/2. Notably, the (155) under mappings of the form
exact scale estimates agree perfectly, whereas the approxi- t0
mate estimates are very similar. In this sense, the two scale t → t0 t vs. t → (156)
t
selection methods have rather similar effects when applied
to a rotationally symmetric Gaussian blob. corresponding to the symmetry

log t0 + log t ↔ log t0 − log t (157)


5.1.1 Theoretical Symmetry Properties Between the Scale
Estimates Since the response properties of the Hessian feature strength
measures D1,norm L and D2,norm L are of similar forms
The similarity between the results of the two scale selection
methods can generally be understood by studying the scale- (1 − 4k) t 2
space signatures that show how the Laplacian and the deter- (D1,norm L)(0, 0; t) = (158)
4π 2 (t0 + t)4
minant of the Hessian responses evolve as function of scale t
(D2,norm L)(0, 0; t) = (159)
2π(t0 + t)2
Table 2 Exact scale estimates obtained from local extrema over scale
vs. weighted averaging over scale for the Laplacian and determinant corresponding symmetry properties follow also for these op-
operators applied to a rotationally symmetric Gaussian blob with scale
erators. These symmetry properties do also extend to mono-
parameter t0 and for a general amount of post-smoothing as determined
by the post-smoothing parameter c tonically increasing transformations ψ of the differential re-
sponses of the form
Operator Extrema over scale Weighted averaging 
 2  
ψ ∇norm L , ψ | det Hnorm L| (160)
∇norm
2 L t0 /(1 + c2 )

t0 /(1 + c2 )

   
det Hnorm t0 / 1 + 2c2 t0 / 1 + 2c2 ψ |D1,norm L| , ψ |D2,norm L| (161)

These symmetry properties do, however, not extend to gen-


Table 3 Approximate scale estimates obtained from local extrema eral values of γ = 1, since such values may lead to a skew-
over scale vs. weighted averaging over scale for the Hessian feature ness in the scale-space signature (see the bottom row in
strength measures I and II applied to a rotationally symmetric Gaus-
Fig. 3).
sian blob with scale parameter t0 and for a specific amount of post-
smoothing with c = 1/2
5.1.2 Calibration Factors for Setting Scale-Invariant
Operator Extrema over scale Weighted averaging Integration Scales
D1,norm L ≈ 0.813 t0 ≈ 0.813 t0
The scale estimates may, however, differ depending on what
D2,norm L ≈ 0.699 t0 ≈ 0.694 t0
differential expression the interest point detector is based on.
J Math Imaging Vis

Fig. 3 Scale-space signatures


computed from a Gaussian blob
with scale parameter t0 = 10 for
(top row) the Laplacian ∇norm
2 L
for γ = 1, (middle row) the
determinant of the Hessian
det Hnorm L for γ = 1 and
(bottom row) the second-order
derivative Lξ ξ for a 1-D signal
and γ = 3/4 so as to give rise to
the peak over scale at tˆ = t0
using (left column) a linear
scaling of the scale parameter t
or (right column) a logarithmic
transformation in terms of
effective scale τ = log t

Hence, if we would like to set an integration scale tint for Table 4 Calibration factors ADL to obtain compensated scale esti-
mates tˆDL ,comp = tˆDL /ADL that lead to tˆDL ,comp = t0 for a rotation-
computing a local image descriptor from the scale estimate ally symmetric Gaussian blob irrespective of the interest point operator
tˆDL , in such a way that the integration scale should be the D L or the post-smoothing parameter c
same
Operator Calibration factor ADL
tint = r 2 t0 (162)
∇norm
2 L 1/(1 + c2 )

for any interest point detector DL applied to a rotationally det Hnorm 1/ 1 + 2c2
symmetric Gaussian blob, irrespective of whether the inter- D1,norm ≈ eθD1 L with θD1 L according to (130)
est points are computed from scale-space extrema or feature D2,norm ≈ eθD2 L with θD2 L according to (135)
trajectories in a scale-space primal sketch, we can parame-
terize the integration scale according to
on local extrema over scale or weighted scale selection lead
tˆDL √
tint = r 2
(163) to a similar and affine covariant scale estimate t1 t2 for the
ADL determinant of the Hessian operator det Hnorm L.
with the calibration factor ADL determined from the results For the Laplacian ∇norm 2 L and the Hessian feature
in Table 4. strength measures D1,norm L and D2,norm L, the scale esti-
mates are, however, not affine covariant. Moreover, the two
5.2 Anisotropic Gaussian Blob scale selection methods may lead to different results. When
performing a Taylor expansion of the scale estimate param-
5.2.1 Taylor Expansions for Non-uniform Scaling Factors eterized in terms of a non-uniform scaling factor s relative
Near s = 1 to a base-line scale t0 , the Taylor expansions around s = 1
did, however, agree in their lowest order terms. In this sense,
From the analysis of the scale selection properties of an the two scale selection approaches have approximately sim-
anisotropic Gaussian blob with scale parameters t1 and t2 in ilar properties for the Gaussian blob model for affine image
Sect. 3.3 and Sect. 4.2, we found that scale selection based deformations near the similarity group.
J Math Imaging Vis

From a comparison between the Taylor expansions for of post-smoothing (c = 0) given by


the scale estimates for the different interest point detectors
in Table 5, we can conclude that after the affine covariant 4s
tˆ∇ 2 L = t0 (164)
determinant of the Hessian det Hnorm L, the scale estimate (1 + s)2
obtained from Hessian feature strength measure D1,norm L tˆdet HL = t0 (165)
has the lowest sensitive to affine image deformations fol-
κ(1−s 2 )2 ((1+s 2 ) log s+1−s 2 )
lowed by the Laplacian ∇norm 2 L and the Hessian feature
tˆD1 L = e 4(1−κ)(1+s 2 ) s 2 log s+(1−s 2 )(κ(1−6s 2 +s 4 )+4s 2 ) t0 (166)
strength measure D2,norm L. Corresponding results hold for
the corresponding signed Hessian feature strength measures   1+s
4s 1−s
D̃1,norm L and D̃2,norm L. tˆD2 L = t0 (167)
(1 + s)2
5.2.2 Graphs of Non-uniform Scaling Dependencies for Figure 4 shows graphs of how the scale estimates depend on
General s ≥ 1 the non-uniform scaling parameter s for scale selection by
weighted averaging over scale. As can be seen from these
From the analysis in Sect. 4.3 it follows from (139), (142),
graphs, the behaviour is qualitatively somewhat different for
(144) and (152) that for an anisotropic Gaussian blob with
the four differential expressions.
scale parameters t1 = s t0 and t2 = t0 /s, the scale estimates
For the determinant of the Hessian det Hnorm L, the scale
for weighted scale selection using the Laplacian ∇norm
2 L, de-
estimate coincides with the geometric average of the scale
terminant of the Hessian det Hnorm L and the Hessian feature
strength measures D1,norm L and D2,norm L are in the absence parameters for any non-singular amount of non-uniform
scaling. For the Laplacian operator ∇norm2 L, the scale esti-
mate tˆ∇ 2 L is lower than the geometric average of the scale
Table 5 Taylor expansions for the scale estimates obtained for an
anisotropic Gaussian blob with scale parameters t1 = s t0 and t2 = t0 /s parameters in the two directions, whereas the scale estimates
around s = 1 (assuming s > 1 for the D2,norm L operator). The table are higher than the geometric average for the Hessian fea-
shows the terms in the Taylor expansion that are common for scale se- ture strength measures D1,norm L and D2,norm L. For moder-
lection based on local extrema over scale and scale selection based on
weighted averaging over scale ate values of s ∈ [1, 4], the scale estimates from the Hessian
feature strength measure D1,norm L, are quite close to the
Operator Common terms in series expansion of scale estimate affine covariant geometric average. For the Hessian feature
strength measure D2,norm L on the other hand, the scale esti-
∇norm
2 L (1 − 14 (s − 1)2 + 14 (s − 1)3 + O((s − 1)4 ))t0
mate increases approximately linearly with the non-uniform
det Hnorm L t0
scaling factor s.
D1,norm L (1 + 21 (s − 1) − 21 (s − 1) + O ((s − 1) ))t0
1 2 1 3 4
These graphs also show that the qualitative behaviour de-
D2,norm L (1 + 2 (s − 1) − 8 (s − 1) + O ((s − 1) ))t0
1 1 2 3
rived for Taylor expansions near s = 1 (Table 5) extend to

Fig. 4 Dependency of the scale


estimates tˆDL on the amount of
non-uniform scaling s ∈ [1, 4]
when performing scale selection
by weighted averaging over
scale for a non-uniform
Gaussian blob with scale
parameters t1 = s t0 and
t2 = t0 /s for (upper left) the
Laplacian ∇norm
2 L, (upper right)
the determinant of the Hessian
det Hnorm L, (lower left) the
Hessian feature strength I
D1,norm L and (lower right) the
Hessian feature strength I
D2,norm L. (Horizontal axis:
Non-uniform scaling factor s.
Vertical axis: Scale estimate tˆDL
in units of t0 )
J Math Imaging Vis

Fig. 5 Sample images from a


dataset with 14 images used for
experiments with repeatability
properties under affine image
deformations. (Image size:
560 × 420 pixels)

non-infinitesimal scaling factors up to at least a factor of the image features were ranked on the scale-normalized re-
four. sponse of the differential operator at the scale-space ex-
tremum. For interest points detected by scale linking, the
5.3 Comparison with Experimental Repeatability image features were ranked on a significance measure ob-
Properties tained by integrating the scale-normalized responses of the
differential operator along each feature trajectory, using the
In this section, we shall compare the above mentioned the- methodology described in [39].
oretical results with experimental results of the repeatabil- To make a judgement of whether two image features A
ity properties of the different interest point detectors under and B detected in two differently transformed images f
affine image transformations. and f should be regarded as belonging to the same fea-
ture or not, we associated a scale dependent circle CA and
5.3.1 Experimental Methodology CB with each feature, with the radius of each circle equal to
the detection scale of the corresponding
√ feature measured in
Figure 5 shows a few examples of images from an image units of the standard deviation σ = t of the Gaussian ker-
data set with 14 images from natural environments. Each nel used for scale-space smoothing to the selected scale, in a
such image was subjected to 10 different types of affine im- similar way as the graphical illustrations of scale dependent
age transformations encompassing: image features in previous sections. Then, each such feature
was transformed to the other image domain, using the affine
– a pure scaling U (s) with scaling factor s = 2,
transformation applied to the image coordinates of the cen-
– a pure rotation R(ϕ) with rotation angle ϕ = π/4, and√
ter of the circle and with the scale value transformed to be
√ scalings N (s) with scaling factors s = 2
4
– non-uniform
proportional to the determinant of the affine transformation
and s = 2, respectively, which are repeated and aver-
matrix, t = (det A) t, resulting in two new circular features
aged over four different orientations respectively
CA and CB . The relative amount of overlap between any
Nϕ0 (s) = R(ϕ0 ) N (s) R(ϕ0 )−1 (168) pair of circles was defined by forming the ratio between the
intersection and the union of the two circles in a similar way
with relative orientations of ϕ0 = 0, π/4, π/2 and 3π/4. as [50] define a corresponding ratio for ellipses

For a locally planar surface patch viewed by a scaled ortho- | (CA , CB )|
graphic projection model, the non-uniform rescalings corre- m(CA , CB ) =  (169)
| (CA , CB )|
spond to the amount of foreshortening that arises with slant
angles equal to 32.8◦ and 45◦ , respectively. In this respect, Matching relations were computed in both directions and
the chosen deformations reflect reasonable requirements of a match was then permitted only if a pair of image fea-
robustness to viewing variations for image-based matching tures maximize this ratio over replacements of either image
and recognition. feature by other image features in the same domain and,
For each one of the resulting 14 × (1 + 10) = 154 im- in addition, the value of this ratio was above a threshold
ages, the 400 most significant interest points were detected. m(CA , CB ) > m0 , where we have chosen m0 = 0.40. Fur-
For interest points detected based on scale-space extrema, thermore, only one match was permitted for each image fea-
J Math Imaging Vis

Table 6 Relative ranking of 10 scale-invariant interest point detectors Table 7 Relative ranking of 10 scale-invariant interest point detectors
based on scale selection from scale-space extrema with regard to their based on scale selection by scale linking and weighted averaging over
repeatability scores under a set of 10 different affine image deforma- scale with regard to their repeatability scores under a set of 10 different
tions applied to each one of the 14 images in the image dataset illus- affine image deformations applied to each one of the 14 images in the
trated in Fig. 5 and the extraction of the 400 most significant interest image dataset illustrated in Fig. 5 and the extraction of the 400 most
points from each image significant interest points from each image

Scale selection from local extrema over scale Scale selection by weighted averaging over scale
Feature detector Type Complementary p (400) Feature detector Type Complementary p (400)

κ̃γ −norm (L) extr – 0.876 D1,norm L link-w – 0.887


D1,norm L extr – 0.868 det Hnorm L link-w D1 L > 0 0.886
det Hnorm L extr D1 L > 0 0.867 D̃2,norm L link-w D1 L > 0 0.880
det Hnorm L extr |D̃1 L| > 0 0.852 det Hnorm L link-w |D̃1 L| > 0 0.878
D̃1,norm L extr – 0.849 κ̃γ −norm (L) link-w – 0.873
∇norm
2 L extr – 0.844 det Hnorm L link-w – 0.871
D̃2,norm L extr D1 L > 0 0.842 D̃1,norm L link-w – 0.866
D2,norm L extr D1 L > 0 0.841 D2,norm L link-w D1 L > 0 0.858
∇norm
2 L extr D1 L > 0 0.839 ∇norm
2 L link-w D1 L > 0 0.856
Harris-Laplace extr – 0.781 Harris-Laplace link-w – 0.855

ture, and matching candidates were evaluated in decreasing strength measures D2,norm L and D̃2,norm L, we have also ap-
order of significance. plied complementary thresholding on either D1,norm L > 0
Finally, given that a total number of Nmatched features or D̃1,norm L > 0, which increases the robustness of the im-
matches have been found from N features detected from the age features and improves the repeatability scores of the
image f and N features from the transformed image f , the interest point detectors. For comparison, we do also show
matching performance was computed as corresponding repeatability scores obtained with the Harris-
Laplace operator. With these variations, a total number 10
Nmatched differential interest point detectors are evaluated. Separate
p= (170)
max(N, N ) evaluations are also performed for scale selection from local
extrema over scale vs. scale selection by scale linking and
The matching performance was computed in both directions weighted averaging over scale.
from f to f as well as from f to f and the average value As can be seen from Table 6, the best repeatability prop-
of these performance measures was reported. erties for the interest point detectors based on scale selec-
The evaluation of the matching score was only performed tion from local extrema over scale are obtained for (i) the
for image features that are within the image domain for both rescaled level curve curvature κ̃γ −norm (L), (ii) the Hessian
images before and after the transformation. Moreover, only feature strength measure D1,norm L and (iii) the determinant
features within corresponding scale ranges were evaluated. of the Hessian det Hnorm L.
In other words, if the scale range for the image f before the From Table 7, we can see that the best repeatability prop-
affine transformation was [tmin , tmax ], then image features erties for the interest point detectors based on scale selec-
were searched for in the transformed image f within the tion using scale linking and weighted averaging over scale
, t ] = [(det A) t
scale range [tmin max min , (det A) tmax ]. In addi- are obtained for (i) the Hessian feature strength measure
tion, features in a narrow scale-dependent frame near the im- D1,norm L, (ii) the determinant of the Hessian det Hnorm L
age boundaries were suppressed, to avoid boundary effects and (iii) the Hessian feature strength measure D̃2,norm L.
from influencing the results. In these experiments, we used The repeatability scores are furthermore generally better
tmin = 4 and tmax = 256. for scale selection based on weighted averaging over scale
compared to scale selection based on local extrema over
5.3.2 Relations Between Experimental Results scale.
and Theoretical Results In comparison with our theoretical analysis, we have pre-
viously shown that the response of the determinant of the
Table 6 and Table 7 show the average repeatability scores Hessian det Hnorm L to an affine Gaussian blob is affine co-
obtained for the 7 different interest point detectors we have variant, for both scale selection based on local extrema over
studied in this work. For the Laplacian ∇norm L, determi- scale (97) and scale selection based on scale linking and
nant of the Hessian det Hnorm L and the Hessian feature weighted averaging over scale (143). For the Hessian feature
J Math Imaging Vis

strength measure D1,norm L, a major contribution to this dif- scale selection from weighted averaging of possibly trans-
ferential expression comes from the affine covariant deter- formed magnitude values along the corresponding feature
minant of the Hessian det Hnorm L, and the deviations from trajectories.
affine covariance are small for both scale selection based Experimental results in [39] show that the Hessian fea-
on local extrema over scale (100) and scale selection by ture strength measure D1,norm L and the determinant of the
weighted averaging over scale (148), provided that the non- Hessian det Hnorm L and are also the two interest point de-
uniform image deformations are not too far from the sim- tectors that give the best repeatability properties under real
ilarity group in the sense that the non-uniform scaling fac- (calibrated) perspective image transformations. Thus, the
tor s used in the Taylor expansions is not too far from 1. two best interest point detectors according to our theoreti-
Specifically, the two interest point detectors that have the cal analysis are also the interest point detectors that have the
best theoretical properties under affine image deformations best properties for real image data.
in the sense of having the smallest correction terms in Ta-
ble 5 are also among the top three interest point detectors
6 Summary and Discussion
for both scale selection based on local extrema over scale
and scale selection based on scale linking and weighted av- We have analyzed the scale selection properties of (i) the
eraging over scale. In this respect, the predictions from our Laplacian operator ∇norm
2 L, (ii) the determinant of the Hes-
theoretical analysis are in very good agreement with the ex- sian det Hnorm L, (iii)–(iv) the new Hessian feature strength
perimental results. measures D1,norm L and D2,norm L and (iv) the rescaled
Somewhat more surprisingly the signed Hessian feature level curve curvature operator κ̃γ −norm (L) when applied
strength measure D̃2,norm L performs very well when com- to a Gaussian prototype blob model and using scale se-
bined with scale selection based on weighted averaging over lection from either (vi) local extrema over scale of scale-
scale. The corresponding unsigned entity D2,norm L does not normalized derivatives or (vii) weighted averaging of scale
perform as well, and more comparable to the Laplacian op- values along feature trajectories over scale. We have also
erator ∇norm
2 L. A possible explanation for this is that keep- analyzed (viii) the influence of a secondary post-smoothing
ing the signs of the principal curvatures in the non-linear step after the computation of possibly non-linear differential
minimum operation improves the ability of this operator invariants and (ix) the sensitivity of the scale estimates to
to distinguish between nearby competing image structures, affine image deformations.
a property that is not captured by the analysis of isolated The analysis shows that the scale estimates from the de-
Gaussian blobs. The repeatability properties of the unsigned terminant of the Hessian det Hnorm L are affine covariant for
version D2,norm L are therefore in closer agreement with the the Gaussian blob model for both scale selection based on
presented analysis. local extrema over scale and scale selection by weighted av-
The rescaled level curve curvature κ̃γ −norm (L) performs eraging over scale. The analysis also shows that the scale es-
comparably very well for scale selection based on local ex- timates from the Laplacian operator ∇norm 2 L and the Hessian
trema over scale, whereas it does not perform as well for feature strength measures D1,norm L and D2,norm L are not
scale selection based on scale linking and weighted aver- affine covariant. Out of the latter three operators, the Hessian
feature strength measure D1,norm L has the lowest sensitivity
aging over scale. For scale selection based on local ex-
to affine image deformations outside the similarity group,
trema over scale, our analysis showed that the deviation
whereas the Hessian feature strength measure D2,norm L has
from affine covariance is comparably low (111) for the
the highest sensitivity. The stronger scale dependency of the
value of γ = 7/8 that we used in our experiments. For this
Hessian feature strength measure D2,norm L can be under-
scale selection method, the experimental results are there-
stood from the fact that it responds to the eigenvalue of the
fore in agreement with our theoretical results. Contrary to Hessian matrix corresponding to the slowest spatial varia-
the other interest point detectors, the repeatability properties tions.
of the rescaled level curve curvature operator κ̃γ −norm (L) Experimental results reported in Sect. 5.3 and [39], show
are, however, not improved by scale linking. A possible that the interest point detectors based on the new Hessian
algorithmic explanation to this could be that the rescaled feature strength measure D1,norm L and the determinant of
level curve curvature operator κ̃γ −norm (L) contains a dif- the Hessian det Hnorm L have significantly better repeata-
ferent type of non-linearity that may cause difficulties for bility properties under affine or perspective image transfor-
the scale linking algorithm. Calculating closed-form expres- mations than the Laplacian ∇norm L or the Hessian feature
sions for the scale estimates obtained by weighed averaging strength measure D2,norm L. Corresponding advantages hold
over scale does also seem harder for this operator. We there- relative to the difference-of-Gaussians (DoG) approxima-
fore leave it as an open problem to investigate if also this in- tion of the Laplacian operator or the Harris-Laplace oper-
terest point detector could be improved by scale linking and ator. Hence, the interest point detectors that have the best
J Math Imaging Vis

theoretical properties under affine deformations of Gaussian When making use of a complementary post-smoothing
blobs do also have the best experimental properties. In this operation to suppress spurious variations in the non-linear
respect, the predictions from this theoretical analysis agree feature responses from the interest point detectors to sim-
with corresponding experimental results. plify the task of scale linking, the influence of this post-
When considering scale selection for a rotationally sym- smoothing operation on the scale estimates may, however,
metric Gaussian blob, it is shown that the scale estimates be different for different interest point detectors. If we as-
obtained by scale selection from local extrema over scale sume that scale calibration can be performed based on the
vs. weighted averaging over scale do for γ = 1 (in the 2-D scale selection properties for Gaussian blobs, we have de-
case) lead to similar results for each one of these four op- rived a set of relative calibration or compensation factors for
erators. This similarity can be explained from a symmetry each one of the five main types of interest point detectors
property of the scale-space signature under inversion trans- studied in this paper.
formations of the scale parameter, which correspond to re- To conclude, the analysis presented in this paper provides
flections along the scale axis after a logarithmic transforma- a theoretical basis for a defining a richer repertoire of mech-
tion of the scale parameter in terms of effective scale. Be- anisms for computing scale-invariant image features and im-
cause of this similarity between the scale estimates obtained age descriptors for a wide range of possible applications in
from the two types of scale selection approaches, we may computer vision. In very recent work [42], these generalized
conclude that no additional scale compensation or scale cal- scale-space interest points have been integrated with local
ibration is needed between scale estimates that are obtained scale-invariant image descriptors and been demonstrated to
from weighted averaging over scale vs. local extrema over lead to highly competitive results for image-based matching
scale (provided that γ = 1). and recognition.
Since the commonly used difference-of-Gaussians oper- As outlined in Appendix A.2, these interest point detec-
tors and the analysis of these can be extended to higher-
ator can be seen as a discrete approximation of the Lapla-
dimensional image data in a rather straightforward manner.
cian operator [41], the analysis of the scale selection prop-
erties for the Laplacian operator also provides a theoreti- Acknowledgements I would like to thank the anonymous reviewers
cal model for analyzing the scale selection properties of the for valuable comments and questions that improved the presentation
difference-of-Gaussian keypoint detector used in the SIFT and Oskar Linde for valuable comments on an early version of the
manuscript.
operator [48]. The above mentioned results concerning the
scale selection properties of the Laplacian operator ∇norm
2 L Open Access This article is distributed under the terms of the Cre-
do also extend to the Harris-Laplace operator [49] for which ative Commons Attribution License which permits any use, distribu-
the spatial selection is performed based on spatial extrema of tion, and reproduction in any medium, provided the original author(s)
and the source are credited.
the Harris measure H , whereas the scale selection properties
are solely determined by the scale selection properties of the
Laplacian ∇norm
2 L. Incorporating the scale selection proper-
Appendix
ties of the determinant of the Hessian det Hnorm L, the re-
sults do also extend to the Harris-detHessian, detmu-Laplace A.1 On the Algorithmic Advantages of Post-smoothing
and detmu-detHessian operators proposed in [39] as well as for Scale Linking
other possible types of hybrid approaches.
For scale estimates that are computed algorithmically In [39] a generalized notion of a scale-space primal sketch
from real-world images in an actual implementation, the for differential descriptors was introduced, where pointwise
robustness of image features that are obtained by scale se- image features at different scales are linked into feature tra-
lection from local extrema over scale or weighted scale se- jectories over scale and bifurcation events between these
lection over scale may, however, differ substantially. Ex- feature trajectories are explicitly registered. Specifically, it
perimental results reported in Sect. 5.3 and [39] show that was experimentally shown that:
weighted scale selection leads to interest points that have – a significance measure obtained by integrating scale-
significantly better repeatability properties under perspec- normalized feature responses along each feature trajec-
tive image deformations compared to interest points com- tory allows for a better ranking of interest points than
puted with scale selection from local extrema over scale. ranking on the magnitude of the responses of scale-space
Theoretically, we have also seen that in several cases, extrema, and
weighted scale selection makes it easier to derive closed- – scale selection by weighted averaging over scale allows
form expressions for the scale estimate than for scale se- for interest point detection with better repeatability prop-
lection based on local extrema over scale. In these respects, erties under affine and perspective image deformations
scale selection by weighted averaging over scale can have compared to scale selection from local extrema over scale
both practical and theoretical advantages. of scale-normalized derivatives.
J Math Imaging Vis

Fig. 6 Illustration of the effect of post-smoothing: The left figure use of post-smoothing can reduce the number of multiple responses to
shows extrema of the differential expression D1,norm L computed from similar structures in the image domain. Observe that this operation is
an image detail at scale t = 4 without post-smoothing. The right figure conceptually different from changing the scale at which derivatives are
shows extrema of the same differential expression with complemen- computed, since the post-smoothing is performed after the computa-
tary post-smoothing using c = 3/8. As can be seen from this figure, the tion of the non-linear differential expression

In the original work on a scale-space primal sketch for inten- explicit detection of grey-level blobs are avoided. In order
sity data (as obtained from zero-order scale-space operations to still be able to make use of the highly useful property
without derivative computations) [27, 31] a scale linking al- that a local extremum can be expected to be unlikely to drift
gorithm was proposed based on (i) the detection of grey- outside the support region of the grey-level blob, a local de-
level blobs at any level of scale and (ii) adaptive scale sam- scent/ascent search is instead initiated at the position corre-
pling refinements. By theoretical analysis of singularities in sponding to the local extrema at the adjacent scale. If the de-
scale-space, the generic types of blob events in scale-space scent/ascent search is initiated within the support region of
were classified [26]. Specifically, the theoretical analysis a grey-level blob, this procedure should therefore proceed
showed that although the local drift velocity in scale-space towards the extremum. This approach therefore avoids the
may momentarily tend to infinity near bifurcation events, it complexity problems that otherwise would occur if match-
can be regarded as unlikely that an extremum point moves ing a large number of image features between adjacent scale
outside the support region of the grey-level blob. For this levels.
reason, the matching of image structures over scales used in When computing non-linear expressions from image data
the scale linking algorithm was based on detecting overlaps at, in particular fine or moderate scales, however, it turns
between the support regions of the grey-level blobs at adja- out that one may get several local extrema of low relative
cent scales, and local refinements were performed if the re- amplitude in relation to neighbouring local extrema, which
lations between partially overlapping grey-level blobs could do not correspond to perceptually relevant image structures.
not be decomposed into the generic types of blob events. Since such local perturbations cannot be expected to be sta-
Experimentally, it was shown that this approach allowed for ble under natural imaging conditions, they cannot be ex-
extraction of intuitively reasonable image structures based pected to be perceptually or algorithmically useful. By per-
on very few assumptions. Due to the adaptive scale refine- forming a small amount of additional Gaussian smoothing
ments and the explicit computation of grey-level blobs, this to the computed non-linear differential expression, a large
type algorithm does, however, not appear suitable for real- number of such spurious responses can be suppressed (see
time implementation on a regular processor. Fig. 6). Thereby, the task of scale linking will be simpli-
In [39], a simplified and more efficient type of scale link- fied for the scale linking algorithm. Due to the suppression
ing algorithm was proposed, where scale refinements and of a large number of irrelevant features, the scale linking
J Math Imaging Vis

algorithm will also run significantly faster. Experimentally, – the rescaled Gaussian curvature
the repeatability properties of the resulting interest points do
also become better. G̃γ −norm (L)
From scale invariance arguments it is natural to let the  
= t 4γ Lz (Lxx Lz − 2Lx Lxz ) + L2x Lzz
relation between the post-smoothing scale tpost and the local  
scale level for computing derivatives t be given by × Lz (Lyy Lz − 2Ly Lyz ) + L2y Lzz

− Lz (−Lx Lyz + Lxy Lz − Lxz Ly )
tpost = c2 t (171)
2 
+ Lx Ly Lzz /L2z (175)
for some fixed value of the relative post-smooth scale pa-
rameter c. In our first work on generalized scale-space in- – the unsigned Hessian feature strength measure I
terest points [39], we used c = 1/2 when linking image
structures over scale into feature trajectories. Specifically, D1,γ −norm L
the experiments reported in Sect. 5.3 were performed using ⎧ 3γ
⎨t (det HL − k trace3 HL)
c = 1/2. In connection with our more recent work on inte-
= if det HL − k trace3 HL > 0 (176)
grating the generalized scale-space interest points with local ⎩
0 otherwise
image descriptors [42], we found that c ≈ 3/8 is a better
choice when matching image descriptors under affine and – the unsigned Hessian feature strength measure II
perspective image deformations. Increasing c above c = 1/2  
decreases the repeatability properties for image deforma- D2,γ −norm L = t γ min |λ1 |, |λ2 |, |λ3 | (177)
tions outside the similarity group, whereas decreasing c be-
where λ1 , λ2 and λ3 denote the eigenvalues of the three-
low c = 1/4 leads to a lower suppression of irrelevant fea-
dimensional Hessian matrix HL, whereas Gγ −norm (L)/
tures which affects the repeatability properties of the result-
(L2x + L2y + L2z )2 denotes the Gaussian curvature7 of a level
ing interest points. Beyond this, we have not tried to opti-
surface.
mize the performance, bearing in mind that one could also
consider other types of algorithms for scale linking, which
Related Work In relation to these 3-D extensions of the 2-
could then lead to different trade-offs in terms of the com-
D interest point detectors considered in this work, a three-
plexity of the local matching step vs. efficiency or accuracy
dimensional extension of the Harris operator
considerations.
H3D = det μ − k trace2 μ (178)
A.2 3-D Generalizations of the Scale-Space Interest Point
Detectors defined from the three-dimensional second-moment matrix

 
μ(x; t, s) = ∇L(x − u; t)
For image data f : R3 → R that are defined over a three- u∈R3
dimensional image domain indexed by the image coordi-  T
nates (x, y, z), there are natural ways to extend the interest × ∇L(x − u; t) g(u; s) du (179)
point detectors considered in this work based on the scale-
has been previously demonstrated to be effective for de-
space representation L : R3 × R+ → R of f generated by
tecting sparse spatio-temporal interest points in video data
convolution with the three-dimensional Gaussian kernel
(Laptev and Lindeberg [23]). The three-dimensional Hes-
1 2 2 2
− x +y2t +z
sian feature strength measure D1,γ −norm L can be seen as
g(x, y, z; t) = e (172) a differential analogue to the spatio-temporal Harris opera-
(2πt)3/2
tor, defined from second-order derivatives of image intensi-
using ties only, which should allow for the computation of more
dense sets of interest points, in analogy with the denser sets
– the Laplacian operator

∇γ2 −norm L = t γ trace HL = t γ (Lxx + Lyy + Lzz ) (173) 7 The motivation for multiplying the Gaussian curvature by a power
of the gradient magnitude in (175) is that the resulting operator should
assume high values when the gradient magnitude and the Gaussian cur-
– the determinant of the Hessian
vature are simultaneously high. More generally, also other powers of
 the gradient magnitude could be considered (204). The current power
det Hγ −norm L = t 3γ Lxx Lyy Lzz + 2Lxy Lxz Lyz of four is chosen because it leads to the simplest calculations, in anal-
 ogy with the multiplication by the gradient magnitude raised to the
− Lxx L2yz − Lyy L2xz − Lzz L2xy (174) power of three for the 2-D rescaled level curve curvature operator (7).
J Math Imaging Vis

of interest points that are obtained from the 2-D Hessian fea- (1 − 27k)t 3γ
(D1,γ −norm L)(0, 0; t) = − √ (185)
ture strength measure D1,γ −norm L compared to the interest 16 2π 9/2 (t + t0 )15/2
points detected by the 2-D Harris operator. For 3-D image

data, the parameter k should be in the interval ]0, 27
1
[. (D2,γ −norm L)(0, 0; t) = √ (186)
Also, the spatio-temporal determinant of the Hessian 2 2π 3/2 (t + t0 )5/2
has been used for detecting spatio-temporal interest points Differentiating these expressions with respect to the scale
(Willems et al. [61]) and been demonstrated to allow for parameter t and setting the derivative to zero gives that the
denser sets of spatio-temporal interest points than the spatio- extremum over scale will for all these interest point detectors
temporal Harris operator. The new 3-D interest point de- be assumed at the same scale
tectors D1,γ −norm L, D2,γ −norm L and G̃γ −norm (L) provide

a way to extend this repertoire of 3-D interest point detec- tˆ = t0 (187)
tors. When applying these operators to spatio-temporal im- 5 − 2γ
age data, they should be combined with a spatio-temporal If we would like these interest point detectors to return a
scale-space representation (Lindeberg [33, 40]; Lindeberg scale estimate corresponding to the scale parameter of a
and Fagerström [45]) that allows for different scale param- Gaussian blob, we should therefore choose
eters over the space vs. time. Moreover, the specific cou-
pling between space and time should be explicitly consid- 5
γ3D = (188)
ered when expressing invariance properties over space-time 4
(Lindeberg et al. [40, 44]). Using γ = 1 would otherwise lead to a lower scale estimate
In the following, we shall analyze the scale-selection
2
properties of these 3-D interest point detectors when applied tˆ3−D,γ =1 = t0 (189)
to volumetric (spatial) image data using the same amount of 3
scale-space smoothing over all spatial dimensions. For scale selection based on weighted averaging over scale
of the scale-normalized Laplacian response (183), the corre-
A.2.1 Scale Selection Properties for a 3-D Gaussian Blob sponding scale estimates are for γ ∈]0, 52 [ given by
∞ 3t γ dt
Consider a single 3-D Gaussian blob √
t=0 2 2π 3/2 (t+t0 )5/2 log t t
τ̂∇ 2 L = ∞ 3t γ dt
1 2 2 +z2
− x +y √
t=0 2 2π 3/2 (t+t0 )5/2 t
f (x, y, z) = g(x, y, z; t0 ) = e 2t0
(180)
(2πt0 )3/2    
5 5
= log t0 + ψ(γ ) − ψ −γ = if γ = = log t0
Due to the semi-group property of the Gaussian kernel 2 4
(190)
g(·, ·, ·; t1 ) ∗ g(·, ·, ·; t2 ) = g(·, ·, ·; t1 + t2 ) (181)
where ψ(u) denotes the digamma function, which is the
the scale-space representation of f obtained by Gaussian logarithmic derivative of the Gamma function ψ(u) =
smoothing is given by Γ (u)/Γ (u).
Similarly, from the explicit expression for the determi-
L(x, y, z; t) = g(x, y, z; t0 + t) nant of the Hessian at the center of the Gaussian blob (184),
1 2 +y 2 +z2
− x 2(t it follows that the weighted scale estimate will for γ ∈]0, 52 [
= e 0 +t) (182)
(2π(t0 + t))3/2 be given by
∞ t 3γ dt

A.2.2 The Pure Second-Order Interest Point Detectors t=0 16 2π 9/2 (t+t0 )15/2 log t t
τ̂det HL = ∞
√ t 3γ dt
By differentiation it follows the scale normalized feature t=0 16 2π 9/2 (t+t0 )15/2 t
strength measure at the center (x, y, z) = (0, 0, 0) of the  
15
Gaussian blob will for the Laplacian (173), the determinant = log t0 + ψ(3γ ) − ψ − 3γ
2
of the Hessian (174) and the Hessian feature strength mea-  
sures I (176) and II (177) be given by 5
= if γ = = log t0 . (191)
4
 2  3t γ
∇γ −norm L (0, 0; t) = − √ (183) Due to the similarity between the explicit expressions for
2 2π 3/2 (t + t0 )5/2
the Hessian feature strength measure I (185) and the deter-
t 3γ minant of the Hessian response (184) as well as the simi-
(det Hγ −norm L)(0, 0; t) = − √ (184)
16 2π 9/2 (t + t0 )15/2 larity between the Hessian feature strength measure II (186)
J Math Imaging Vis

and the Laplacian response (185) at the center of a Gaussian which with spherical coordinates x 2 + y 2 + z2 = R 2 be-
blob, the scale estimates for D1,γ −norm L and D2,γ −norm L comes
will be analogous: 2
− 2R
  R 2 t 3γ e t+t0
15 G̃γ −norm (L) = (199)
τ̂D1 L = log t0 + ψ(3γ ) − ψ − 3γ 64 π 6 (t + t0 )10
2
  This expression assumes its maximum value over R when
5
= if γ = = log t0 (192) 
4 t0 + t
  R= (200)
5 2
τ̂D2 L = log t0 + ψ(γ ) − ψ −γ
2 for which the rescaled surface curvature assumes the value
 
5
= if γ = = log t0 (193) t 3γ
4 G̃γ −norm (L) = (201)
128 e π 6 (t + t0 )9
When expressed in terms of the regular scale parameter
Setting the derivative of this expression with respect to the
scale parameter to zero gives
tˆ = e τ̂
(194)
γ
tˆ = t0 (202)
the weighted scale selection method does hence for a rota- 3−γ
tionally symmetric Gaussian blob lead to similar scale esti-
mates as are obtained from local extrema over scale of γ - for which the γ -normalized magnitude response is
normalized derivatives (29) when γ = 5/4. For other values  3γ
(3 − γ )9 γ t0
of γ , the scale estimates are biased G̃γ −norm (L) = (203)
2519424 e π 6 t09 3−γ
tˆ = C t0 (195)
A corresponding analysis can be carried out for the Gaus-
sian curvature modulated by other powers of the gradient
by a scale factor equal to
magnitude
5
C1 = eψ(γ )−ψ( 2 −γ ) (196) G̃γ −norm (L)
(a)
G̃γ −norm (L) = (204)
(L2x + L2y + L2z )a
for the Laplacian ∇norm
2
and the Hessian feature strength
L
measure D2,norm L operators, and a bias factor where a < 2. Then, the corresponding scale estimate be-
15
comes
C2 = eψ(3γ )−ψ( 2 −3γ ) (197)
(3 − a) γ
tˆ = t0 (205)
for the determinant of the Hessian det Hnorm L and the Hes- 9 − 4a − (3 − a) γ
sian feature strength measure D1,norm L operators. For general values of a, the corresponding intermediate re-
By following the methodology outlined in Sects. 3.2, 3.3, sults are, however, more complex.
4.2 and 4.3, a corresponding more detailed analysis can be
performed concerning the influence of the post-smoothing A.3 Interpreting the Parameter γ in Terms
operation and affine image deformations for these 3-D inter- of the Dimensionality of the Image Features
est point detectors.
The value γ = 5/4 obtained from the analysis of a 3-D
A.2.3 The Rescaled Gaussian Curvature Operator Gaussian blob in Appendix A.2.2 (Eq. (188)) can be com-
pared to the values obtained by requiring the purely second-
When computing the rescaled Gaussian curvature order differential entities to respond to a 2-D Gaussian blob
G̃γ −norm (L) for the scale-space representation of a three- at t = t0 (Eq. (29))
dimensional spherically symmetric Gaussian blob with scale
parameter t0 , we obtain γ2D = 1 (206)

2 +y 2 +z2 ) or the value of the γ parameter obtained by requiring the


− 2(x
t 3γ (x 2 + y 2 + z2 )e t+t0
second-order derivative operator to respond to a 1-D Gaus-
G̃γ −norm (L) = (198)
64π 6 (t + t0 )10 sian blob (alternatively a 1-D Gaussian ridge embedded in a
J Math Imaging Vis

higher-dimensional space) at t = t0 [34] References


3
γ1D = (207) 1. Baumberg, A.: Reliable feature matching across widely separated
4 views. In: Proc. CVPR, Hilton Head, SC, pp. 1774–1781 (2000)
2. Bay, H., Ess, A., Tuytelaars, T., van Gool: Speeded up robust
More generally, for a D-dimensional Gaussian intensity pro-
features (SURF). Comput. Vis. Image Underst. 110(3), 346–359
file (2008)
D
3. Bevensee, R.: Maximum Entropy Solutions to Scientific Prob-
2
1 − i=1 xi lems. Prentice Hall, New York (1993)
f (x1 , . . . , xD ) = g(x1 , . . . , xD ; t0 ) = D/2
e 2t0
4. Bretzner, L., Lindeberg, T.: Feature tracking with automatic selec-
(2πt0 ) tion of spatial scales. Comput. Vis. Image Underst. 71(3), 385–392
(208) (1998)
5. Chomat, O., de Verdiere, V., Hall, D., Crowley, J.: Local scale
embedded in N ≥ D dimensions, the scale-space represen- selection for Gaussian based description techniques. In: Proc.
ECCV’00, Dublin, Ireland. Lecture Notes in Computer Science,
tation is given by
vol. 1842, pp. 117–133. Springer, Berlin (2000)
6. Constable, R.T., Henkelman, R.M.: Why MEM does not work
L(x1 , . . . , xN ; t) = g(x1 , . . . , xD ; t0 + t) in MR image reconstruction. Magn. Reson. Med. 14(1), 12–25
D
i=1 xi
2 (1990)
1 −
= e 2(t0 +t)
(209) 7. Danielsson, P.E., Lin, Q., Ye, Q.Z.: Efficient detection of second-
(2π(t0 + t))D/2 degree variations in 2D and 3D images. J. Vis. Commun. Image
Represent. 12(3), 255–305 (2001)
with the corresponding γ -normalized Laplacian response 8. Donoho, D.L., Johnstone, I.M., Hoch, J., Stern, A.S.: Maximum
 2  entropy and the nearly black object. J. R. Stat. Soc., Ser. B
∇γ −norm L (x1 , . . . , xN ; t) (Methodological) 54, 41–81 (1992)
9. Elder, J., Zucker, S.: Local scale control for edge detection and

D
(x 2 − t0 − t) blur estimation. IEEE Trans. Pattern Anal. Mach. Intell. 20(7),
=t γ i
g(x1 , . . . , xD ; t0 + t) (210) 699–716 (1998)
(t0 + t)2
i=1 10. Florack, L.M.J.: Image Structure. Series in Mathematical Imaging
and Vision. Springer, Berlin (1997)
which assumes the following value at the origin 11. Frangi, A.F., Niessen, W.J., Hoogeveen, R.M., vanWalsum, T.,
 2  Viergever, M.A.: Model-based quantitation of 3D magnetic res-
D tγ
∇γ −norm L (0, . . . , 0; t) = − (211) onance angiographic images. IEEE Trans. Med. Imaging 18(10),
(2π(t0 + t))1+D/2 946–956 (1999)
12. Gårding, J., Lindeberg, T.: Direct computation of shape cues us-
Differentiating this expression with respect to t gives and ing scale-adapted spatial derivative operators. Int. J. Comput. Vis.
setting the derivative to zero gives 17(2), 163–191 (1996)
13. Griffin, L.D.: The second order local-image-structure solid. IEEE

tˆ = t0 (212) Trans. Pattern Anal. Mach. Intell. 29(8), 1355–1366 (2007)
D + 2(1 − γ ) 14. ter Haar Romeny, B.: Front-End Vision and Multi-Scale Image
Analysis. Springer, Berlin (2003)
Requiring this scale estimate to be equal to t0 implies that 15. Hall, D., de Verdiere, V., Crowley, J.: Object recognition us-
the γ -value for a pure second-order operator should there- ing coloured receptive fields. In: Proc. ECCV’00, Dublin, Ire-
fore be related to the dimensionality D of the image features land. Lecture Notes in Computer Science, vol. 1842, pp. 164–177.
Springer, Berlin (2000)
they should respond to according to 16. Harris, C., Stephens, M.: A combined corner and edge detector.
1 D In: Alvey Vision Conference, pp. 147–152 (1988)
γ= + (213) 17. Iijima, T.: Observation theory of two-dimensional visual patterns.
2 4 Tech. rep., Papers of Technical Group on Automata and Automatic
Control, IECE, Japan (1962)
Note that the 2-D case is special in the sense that only for
18. Kadir, T., Brady, M.: Saliency, scale and image description. Int. J.
this dimensionality D = 2 does scale selection based on the Comput. Vis. 45(2), 83–105 (2001)
most scale-invariant choice γ = 1 lead to scale estimates 19. Kang, Y., Morooka, K., Nagahashi, H.: Scale invariant texture
that are equal to the diffuseness parameter t0 of a Gaussian analysis using multi-scale local autocorrelation features. In: Proc.
intensity profile. Scale Space and PDE Methods in Computer Vision (Scale-
Space’05). Lecture Notes in Computer Science, vol. 3459, pp.
It can also be noted that only for two-dimensional blobs
363–373. Springer, Berlin (2005)
will the corresponding γ -normalized magnitude values of 20. Koenderink, J.J.: The structure of images. Biol. Cybern. 50, 363–
scale-space extrema be independent of the size of the blob, 370 (1984)
whereas in other dimensionalities the corresponding mag- 21. Koenderink, J.J., van Doorn, A.J.: Generic neighborhood oper-
nitude values need to be normalized by a scale-dependent ators. IEEE Trans. Pattern Anal. Mach. Intell. 14(6), 597–605
(1992)
correction factor tˆ1−γ to lead to scale-invariant magnitude 22. Krissian, K., Malandain, G., Ayache, N., Vaillant, R., Trousset, Y.:
values that are independent of the diffuseness t0 of the D- Model-based detection of tubular structures in 3D images. Com-
dimensional Gaussian intensity profile. put. Vis. Image Underst. 80(2), 130–171 (2000)
J Math Imaging Vis

23. Laptev, I., Lindeberg, T.: Space-time interest points. In: Proc. 9th 42. Lindeberg, T.: Distinctiveness and matching properties of general-
Int. Conf. on Computer Vision, Nice, France, pp. 432–439 (2003) ized scale-space interest points (2012). Unpublished manuscript
24. Lazebnik, S., Schmid, C., Ponce, J.: A sparse texture representa- 43. Lindeberg, T.: Scale Selection. Encyclopedia of Computer Vision.
tion using local affine regions. IEEE Trans. Pattern Anal. Mach. Springer, Berlin (2012, in press)
Intell. 27(8), 1265–1278 (2005) 44. Lindeberg, T., Akbarzadeh, A., Laptev, I.: Galilean-corrected
25. Li, Y., Tax, D.M.J., Loog, M.: Supervised scale-invariant segmen- spatio-temporal interest operators. In: International Conference on
tation (and detection). In: Proc. Scale Space and Variational Meth- Pattern Recognition, Cambridge, pp. I:57–62 (2004)
ods in Computer Vision (Scale-Space’11), Ein Gedi, Israel. Lec-
45. Lindeberg, T., Fagerström, D.: Scale-space with causal time direc-
ture Notes in Computer Science, vol. 6667, pp. 350–361. Springer,
tion. In: Proc. ECCV’96, Cambridge, UK, vol. 1064, pp. 229–240.
Berlin (2012)
Springer, Berlin (1996)
26. Lindeberg, T.: Scale-space behaviour of local extrema and blobs.
J. Math. Imaging Vis. 1(1), 65–99 (1992) 46. Lindeberg, T., Gårding, J.: Shape-adapted smoothing in estimation
of 3-D depth cues from affine distortions of local 2-D structure.
27. Lindeberg, T.: Detecting salient blob-like image structures and
Image Vis. Comput. 15, 415–434 (1997)
their scales with a scale-space primal sketch: a method for focus-
of-attention. Int. J. Comput. Vis. 11(3), 283–318 (1993) 47. Loog, M., Li, Y., Tax, D.: Maximum membership scale selection.
28. Lindeberg, T.: Effective scale: a natural unit for measuring scale- In: Multiple Classifier Systems. Lecture Notes in Computer Sci-
space lifetime. IEEE Trans. Pattern Anal. Mach. Intell. 15(10), ence, vol. 5519, pp. 468–477. Springer, Berlin (2009)
1068–1074 (1993) 48. Lowe, D.: Distinctive image features from scale-invariant key-
29. Lindeberg, T.: On scale selection for differential operators. In: points. Int. J. Comput. Vis. 60(2), 91–110 (2004)
Høgdra, K.H.K.A., Braathen, B. (eds.) Proc. 8th Scandinavian 49. Mikolajczyk, K., Schmid, C.: Scale and affine invariant interest
Conf. on Image Analysis, pp. 857–866. Norwegian Society for point detectors. Int. J. Comput. Vis. 60(1), 63–86 (2004)
Image Processing and Pattern Recognition, Tromsø (1993) 50. Mikolajczyk, K., Tuytelaars, T., Schmid, C., Zisserman, A.,
30. Lindeberg, T.: Scale-space theory: a basic tool for analysing Matas, J., Schaffalitzky, F., Kadir, T., van Gool, L.: A compari-
structures at different scales. J. Appl. Stat. 21(2), 225–270 (1994). son of affine region detectors. Int. J. Comput. Vis. 65(1–2), 43–72
Also available from. http://www.csc.kth.se/~tony/abstracts/ (2005)
Lin94-SI-abstract.html 51. Mrázek, P., Navara, M.: Selection of optimal stopping time for
31. Lindeberg, T.: Scale-Space Theory in Computer Vision. The nonlinear diffusion filtering. Int. J. Comput. Vis. 52(2–3), 189–
Kluwer International Series in Engineering and Computer Sci- 203 (2003)
ence. Springer, Berlin (1994)
52. Negre, A., Braillon, C., Crowley, J.L., Laugier, C.: Real-time time-
32. Lindeberg, T.: Direct estimation of affine deformations of bright-
to-collision from variation of intrinsic scale. Exp. Robot. 39, 75–
ness patterns using visual front-end operators with automatic scale
84 (2008)
selection. In: Proc. 5th Int. Conf. on Computer Vision, Cambridge,
MA, pp. 134–141 (1995) 53. Pedersen, K.S., Nielsen, M.: The Hausdorff dimension and scale-
33. Lindeberg, T.: Linear spatio-temporal scale-space. In: ter Haar space normalisation of natural images. J. Math. Imaging Vis.
Romeny, B.M., Florack, L.M.J., Koenderink, J.J., Viergever, M.A. 11(2), 266–277 (2000)
(eds.) Scale-Space Theory in Computer Vision: Proc. First Int. 54. Sato, Y., Nakajima, S., Shiraga, N., Atsumi, H., Yoshida, S.,
Conf. Scale-Space’97, Utrecht, The Netherlands. Lecture Notes Koller, T., Gerig, G., Kikinis, R.: 3D multi-scale line filter for seg-
in Computer Science, vol. 1252, pp. 113–127. Springer, Berlin mentation and visualization of curvilinear structures in medical
(1997). Extended version available as technical report ISRN KTH images. Med. Image Anal. 2(2), 143–168 (1998)
NA/P–01/22–SE from KTH 55. Shi, J., Tomasi, C.: Good features to track. In: Proc. CVPR, pp.
34. Lindeberg, T.: Edge detection and ridge detection with automatic 593–600 (1994)
scale selection. Int. J. Comput. Vis. 30(2), 117–154 (1998) 56. Sporring, J., Colios, C.J., Trahanias, P.E.: Generalized scale-
35. Lindeberg, T.: Feature detection with automatic scale selection. selection. In: Proc. International Conference on Image Processing
Int. J. Comput. Vis. 30(2), 77–116 (1998) (ICIP’00), Vancouver, Canada, pp. 920–923 (2000)
36. Lindeberg, T.: A scale selection principle for estimating image de- 57. Tuytelaars, T., van Gool, L.: Matching widely separated views
formations. Image Vis. Comput. 16(14), 961–977 (1998) based on affine invariant regions. Int. J. Comput. Vis. 59(1), 61–85
37. Lindeberg, T.: Principles for automatic scale selection. In: (2004)
Handbook on Computer Vision and Applications, pp. 239– 58. Tuytelaars, T., Mikolajczyk, K.: A survey on local invariant fea-
274. Academic Press, Boston (1999). Also available from tures. In: Foundations and Trends in Computer Graphics and Vi-
http://www.csc.kth.se/cvap/abstracts/cvap222.html sion, vol. 3(3). Now Publishers, Boston (2008)
38. Lindeberg, T.: Scale-space. In: Wah, B. (ed.) Encyclopedia of
59. Uffink, J.: Can the maximum entropy principle be explained as
Computer Science and Engineering, pp. 2495–2504. Wiley,
a consistency requirement? Stud. Hist. Philos. Mod. Phys. 26(3),
Hoboken (2008). doi:10.1002/9780470050118.ecse609. Also
223–261 (1995)
available from http://www.nada.kth.se/~tony/abstracts/Lin08-
EncCompSci.html 60. Weickert, J., Ishikawa, S., Imiya, A.: Linear scale-space has first
39. Lindeberg, T.: Generalized scale-space interest points: scale-space been proposed in Japan. J. Math. Imaging Vis. 10(3), 237–252
primal sketch for differential descriptors. Int. J. Comput. Vis. (1999)
(2010) (original version submitted in June 2010) 61. Willems, G., Tuytelaars, T., van Gool, L.: An efficient dense and
40. Lindeberg, T.: Generalized Gaussian scale-space axiomatics com- scale-invariant spatio-temporal interest point detector. In: Proc.
prising linear scale-space, affine scale-space and spatio-temporal ECCV’08, Marseille, France. Lecture Notes in Computer Science,
scale-space. J. Math. Imaging Vis. 40(1), 36–81 (2011) vol. 5303, pp. 650–663. Springer, Berlin (2008)
41. Lindeberg, T.: Scale invariant feature transform. Scholarpedia 62. Witkin, A.P.: Scale-space filtering. In: Proc. 8th Int. Joint Conf.
7(5), 10491 (2012) Art. Intell, Karlsruhe, Germany, pp. 1019–1022 (1983)
J Math Imaging Vis

Tony Lindeberg is a professor in nition, focus-of-attention and shape. He has developed theories and
Computer Science at KTH Royal methodologies for continuous and discrete scale-space representation,
Institute of Technology in Stock- detection of salient image structures, automatic scale selection, scale-
holm, Sweden. He was born in invariant image features, affine invariant features, affine and Galilean
Stockholm in 1964, received his normalization, temporal and spatio-temporal scale-space concepts as
M.S.c. degree in 1987, his Ph.D. well as spatial and spatio-temporal image descriptors for image-based
degree in 1991, became docent in recognition. He does also work on computational modelling of biologi-
1996, and was appointed professor cal vision and has previously worked on applications in medical image
in 2000. He was a Research Fellow analysis and gesture recognition. He is author of the book Scale-Space
at the Royal Swedish Academy of Theory in Computer Vision.
Sciences between 2000 and 2010.
His research interests in computer
vision relate to scale-space repre-
sentation, image features, object
recognition, spatio-temporal recog-

Vous aimerez peut-être aussi