Computer Vision, 2007. ICCV 2007. IEEE 11th International Conference on

Learning Globally-Consistent Local Distance Functions for Shape-Based

Image Retrieval and Classification

Computer Vision, 2007. ICCV 2007. IEEE 11th International Conference on

Andrea Frome , EECS, UC BerkeleyYoram Singer, Google, Inc

Fei Sha , EECS, UC BerkeleyJitendra Malik, EECS, UC Berkeley

Outline

• Introduction• Training step• Testing step• Experiment & Result• Conclusion

Outline

• Introduction• Training step• Testing step• Experiment & Result• Conclusion

What we do?

• Goal– classify an image to a more appropriate category

• Machine learning• Two steps– Training step– Testing step

Outline

• Introduction

• Training step• Testing step• Experiment & Result• Conclusion

Flow chart: training

Generate features each image from dataset, ex: SIFT or geometric blur

Input distances to SVM for training , evaluate W

Compute distance dji, dki

Flow chart: training

Generate features each image from dataset, ex: SIFT or geometric blur

Input distances to SVM for training , evaluate W

Compute distance dji, dki

Choosing features

• Dataset: Caltech101• Patch-based Features– SIFT• Old school

– Geometric Blur• It’s a notion of blurring• The measure of similarity between image patches• The extension of Gaussian blur

Geometric blur

Flow chart: training

Generate features each image from dataset, ex: SIFT or geometric blur

Input distances to SVM for training , evaluate W

Compute distance dji, dki

Triplet

• dji is the distance from image j to i• It’s not symmetric, ex: dji ≠ dij• dki > dji

dji dki

How to compute distance

• L2 norm

12

3

dji, 1

m features

dji, 1distance vector dji

Image j

Image i

Example

• Given 101 category, 15 images each category101*15

Feature j

101*15

distance vector

distance vector

Image j vs training data

Flow chart: training

Generate features each image from dataset, ex: SIFT or geometric blur

Input distances to SVM for training , evaluate W

Compute distance dji, dki

Machine learning: SVM

• Support Vector Machine• Function: Classify prediction• Supervised learning• Training data are n dimension vector

Example

• Male investigate– Annual income– Free time

• Have girlfriend?

Ex: Training data

spacefree

income

vector

Mathematical expression(1/2)

Mathematical expression(2/2)

Support vector

Model

free

income

But the world is not so ideal.

Real world data

Hyper-dimension

Error cut

SVM standard mathematical expression

Trade-off

In this paper

• Goal: to get the weight vector W

101*15

feature

Image weight wj of W wj, 1

wj

Visualization of the weights

How to choose Triplets?

• Reference Image– Good friend - In the same class– Bad friend - In the different class

• Ex: 101category, 15 images per category– 14 good friends & 15*100(1500) bad friends– 15*101(1515) reference images– total of about 31.8 million triplets

Mathematical expression(1/2)

• Idealistic: • Scaling:• Different:

The length of Weight i

0 0

triplet

Mathematical expression(2/2)

• Empirical loss:

• Vector machine:

Dual problem

•

•

Dual variable

• Iterate the dual variables:

Early stopping

• Satisfy KTT condition– In mathematics, a solution in

nonlinear programming to be optimal.• Threshold– Dual variable update falls below a value

Outline

• Introduction• Training step

• Testing step• Experiment & Result• Conclusion

Flow chart: testing

Query an image i

Output the most appropriate category

Calculate Dxi, x is all training data, except itself.

Flow chart: testing

Query an image i

Output the most appropriate category

Calculate Dxi, x is all training data, except itself.

Query image?

• Goal: classify the query image to an appropriate class

• Using the remaining images in the dataset as the query image

Flow chart: testing

Query an image i

Output the most appropriate category

Calculate Dxi, x is all training data, except itself.

Distance function(1/2)

• Query image i

Image i feature

101*15

distance vector

distance vector

Image i vs all training data

dxi, 1

Distance function(2/2)

•

101*15

Image I vs all the training data

Dji

Flow chart: testing

Query an image i

Output the most appropriate category

Calculate Dxi, x is all training data, except itself.

How to choose the best image?

• Modified 3-NN classifier• no two images agree on the class within the

top 10– Take the class of the top-ranked image of the 10

Outline

• Introduction• Training step• Testing step

• Experiment & Result• Conclusion

Experiment & Result

• Caltech 101• Feature– Geometric blur (shape feature)– HSV histograms (color feature)

• 5, 10, 15, 20 training images per category

Confusion matrix for 15

Outline

• Introduction• Training step• Testing step• Experiment & Result

• Conclusion

Conclusion

• Learning Globally-Consistent Local Distance Functions for Shape-Based Image Retrieval and Classification