Effective Image Database Searc

h via Dimensionality Reduction

Anders Bjorholm Dahl and Henrik Aans

IEEE Computer Society Conference on
Computer Vision and Pattern Recognition
Workshops
Outline
Introduction
Methods
LF-clustering
Experiments and Results

Discussion and Conclusion

Introduction
The bag-of-words approach
1. Feature extraction from the database image
s
2. Building the bag-of-words representation
3. Searching with a query image
Introduction
The Bag-of-word Model
Methods
Feature representation
Clustering
Feature assignment

Image matching
Feature representation
PCA is applied to reduce the dimensionali
ty of the feature vectors

The reduction of the SIFT descriptor is fro

m 128 to between 3 and 12 dimensions
After dimension reduction we add color t
o our features
the mean RGB value in a 10 10 pixels patch
around the localization of each feature
Feature representation
s [ sPCA , (1 ) sRGB ]
sPCA is the PCA reduced SIFT feature
sRGB is the mean RGB values

is a weighing parameter( 0.5 )

1. sPCA , sRGB normalized to unit length
2. normalized s
Clustering
Similar but faster than Mean-shift clusteri
ng
Feature assignment
Similarity of images are found by compari
ng frequency vectors of a query image to i
mages in the database

Give each visual words a weight[16]

N

wi log( )
ni
wi : the weight of word i
N : the total number of images in the database
ni : the number of images where word i occurs
[16] D. Nister and H. Stewenius. Scalable recognition with a vocabulary tree.
In IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
volume 2, pages 21612168, June 2006.
Image matching
Frequency vectors are compared using th
eL1
norm
which is found to be superior to the euclidea
L1 n distance[16]
norm gives equal weight to the overlap
ping and non-overlapping parts
Inverted files are used for fast image retri
eval
Experiments and Results
Data set
first 1400 images form [16]
a series of 4 images of the same scene
Use three of the images from one scene to tr
ain the model and the last for testing
The test result is the percentage of the correc
t images ranked in top 3
data set is relatively small
http://www.vis.uky.edu/~stewe/ukbench/
Experiments and Results
Data set
 Experiments and Results
Experiments
Color added PCA SIFT
3, 8, and 12 dimensional PCA SIFT features
added features are 6, 11, and 15 dimensions
compare with SIFT features reduced with PCA to
6, 11 and 15 dimensions (without color)
Clustering experiments
LF-clustering
from 8,000 to 12,000 clusters
k-means

10 clusters in 4 levels resulting in 10,000 clusters
 Experiments and Results
Results
 Experiments and Results
Results
 Discussion and Conclusion
did not apply LF-clustering to the 128 dim
ensional SIFT features, because it perfor
med very poorly

for future work the model should be teste

d on a larger set of data
A problem of the design of the bag-of-wor
ds model is it static nature
not designed for adding or removing images f
rom the database