29
Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

Embed Size (px)

Citation preview

Page 1: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

Boris Babenko, Steve Branson, Serge BelongieUniversity of California, San Diego

ICCV 2009, Kyoto, Japan

Page 2: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Recognizing multiple categories– Need meaningful similarity metric / feature space

Page 3: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Recognizing multiple categories– Need meaningful similarity metric / feature space

• Idea: use training data to learn metric, plug into kNN– Goes by many names:

• metric learning• cue combination/weighting• kernel combination/learning• feature selection

Page 4: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Learn a single global similarity metric

Labeled Dataset

Mon

olith

ic

Query Image Similarity Metric

Cate

gory

4Ca

tego

ry 3

Cate

gory

2Ca

tego

ry 1

[ Jones et al. ‘03,Chopra et al. ‘05,Goldberger et al. ‘05,Shakhnarovich et al. ‘05Torralba et al. ‘08]

Page 5: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Learn similarity metric for each category (1-vs-all)

Labeled Dataset

Mon

olith

icCa

tego

rySp

ecifi

c

Query Image Similarity Metric

Cate

gory

4Ca

tego

ry 3

Cate

gory

2Ca

tego

ry 1

[ Varma et al. ‘07,Frome et al. ‘07,Weinberger et al. ‘08Nilsback et al. ’08]

Page 6: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Per category:– More powerful– Do we really need thousands of metrics?– Have to train for new categories

• Global/Monolithic:– Less powerful– Can generalize to new categories

Page 7: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Would like to explore space between two extremes

• Idea: – Group categories together – Learn a few similarity metrics, one for each super-

category

Page 8: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Learn a few good similarity metrics

Labeled Dataset

Mon

olith

icM

uSL

Cate

gory

Spec

ific

Query Image Similarity Metric

Cate

gory

4Ca

tego

ry 3

Cate

gory

2Ca

tego

ry 1

Page 9: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Need some framework to work with…• Boosting has many advantages:

– Feature selection– Easy implementation– Performs well

• Can treat metric learning as binary classification

Page 10: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Training data:

• Generate pairs:– Sample negative pairs

( , ), 0

Images

Category Labels

( , ), 1

Page 11: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Train similarity metric/classifier:

Page 12: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Choose to be binary -- i.e.• = L1 distance over binary vectors

– Can pre-compute for training data – Efficient to compute (XOR and sum)

• For convenience:

[Shakhnarovich et al. ’05, Fergus et al. ‘08]

Page 13: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Given some objective function• Boosting = gradient ascent in function space• Gradient = example weights for boosting

chosen weak classifier

other weak classifiers

function space

current strong classifier

[Friedman ’01, Mason et al. ‘00]

Page 14: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Goal: train and recover mapping• At runtime

– To compute similarity of query image touse

Cate

gory

4Cat

egor

y 3

Cate

gory

2Ca

tego

ry 1

Page 15: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Run pre-processing to group categories (i.e. k-means), then train as usual

• Drawbacks:– Hacky / not elegant– Not optimal: pre-processing not informed by class

confusions, etc.

• How can we train & group simultaneously?

Page 16: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Definitions:

Sigmoid Function Parameter

Page 17: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Definitions:

Page 18: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Definitions:

How well works with category

Page 19: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Objective function:

• Each category “assigned” to classifier

Page 20: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Replace max with differentiable approx.

where is a scalar parameter

Page 21: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Each training pair has weights

Page 22: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Intuition:

Approximation of Difficulty of pair(like regular boosting)

Page 23: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

10 20 30

0.5

1

1.5

2

x 10-4

10 20 30

1

2

3

4

5

6x 10

-5

w1i

w2i

w3i

(boosting iteration) (boosting iteration)

Difficult PairAssigned to

Easy PairAssigned to

Page 24: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan
Page 25: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Created dataset with hierarchical structure of categories

5 10 15 200.65

0.7

0.75

0.8

K (number of classifiers)

Acc

ura

cy

MuSL+retrainMuSLk-meansRandMonolithicPer Cat

Merged categories from:• Caltech 101 [Griffin et al.]• Oxford Flowers [Nilsback et al.]• UIUC Textures [Lazebnik et al.]

Page 26: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

MuS

L

k-m

eans

Page 27: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

Training more metrics overfits!

New categories only Both new and old categories

Page 28: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Studied categorization performance vs number of learned metrics

• Presented boosting algorithm to simultaneously group categories and train metrics

• Observed overfitting behavior for novel categories

Page 29: Boris Babenko, Steve Branson, Serge Belongie University of California, San Diego ICCV 2009, Kyoto, Japan

• Supported by– NSF CAREER Grant #0448615 – NSF IGERT Grant DGE-0333451– ONR MURI Grant #N00014-08-1-0638– UCSD FWGrid Project (NSF Infrastructure Grant

no. EIA-0303622)