20
Knowing a Good HOG Filter When You See It: Efficient Selection of Filters for Detection Ejaz Ahmed 1 , Gregory Shakhnarovich 2 , and Subhransu Maji 3 1 University of Maryland, College Park 2 Toyota Technological Institute at Chicago 3 University of Massachusetts, Amherst

Visual Category as Collection of filters

  • Upload
    travis

  • View
    21

  • Download
    1

Embed Size (px)

DESCRIPTION

Knowing a Good HOG Filter W hen Y ou S ee I t: Efficient Selection of Filters for Detection Ejaz Ahmed 1 , Gregory Shakhnarovich 2 , and Subhransu Maji 3 1 University of Maryland, College Park 2 Toyota Technological Institute at Chicago 3 University of Massachusetts, Amherst. - PowerPoint PPT Presentation

Citation preview

Page 1: Visual Category as Collection of filters

Knowing a Good HOG Filter When You See It:

Efficient Selection of Filters for Detection

Ejaz Ahmed1, Gregory Shakhnarovich2, and Subhransu Maji3

1 University of Maryland, College Park

2 Toyota Technological Institute at Chicago

3 University of Massachusetts, Amherst

Page 2: Visual Category as Collection of filters

Visual Category as Collection of fi lters

Poselets Mid Level Discriminative Patches

Exemplar SVMs

Page 3: Visual Category as Collection of filters

Candidate Generation

Candidate Generation

Pool of Filters

Filters are generated using positives and negatives examples.

Generation of a large pool of fi lters.

Positives Negatives

Filter (SVM Classifier)

Page 4: Visual Category as Collection of filters

Two sources of ineffi ciency

Candidate Selection

CandidateSelection

Selected Filters (n)

(n << N)

Pool of Filters (N)

Impractical to use all generated fi lters.

Redundancy

Noise

good

bad

Page 5: Visual Category as Collection of filters

Candidate Selection Cont…

Expensive Evaluation

Selected Filters (n)

(n << N)

Run as detectorPool of Filters (N)

Bottleneck

Page 6: Visual Category as Collection of filters

What we Propose

Expensive Evaluation

Selected Filters (n)

(n << N)

Run as detectorPool of Filters (N)

By passes explicit evaluation

fast

Our Contribution : fast automatic selection of a subset of discriminative and non redundant filters given a collection of filters

Page 7: Visual Category as Collection of filters

Category Independent Model

Images +/-

Pool (Candidate Filters)N >> n

Selected Filters (n)

fast slow

(w , λ)

Test Category

fast

N candidates

n selected

fast

Can rank fi lters as accurately as a direct evaluation on thousands of examples.

Page 8: Visual Category as Collection of filters

Poselets

Poselets are semantically aligned discriminative patterns that capture parts of object.

Patches are often far visually, but they are close semantically

Page 9: Visual Category as Collection of filters

Poselet Architecture

Candidate Selection :

Candidate Generation :

24

76

Timings

generation

selection

Total Time = 20hrs

Page 10: Visual Category as Collection of filters

Save significantly in training time if we can quickly select small set of relevant exemplars.

ESVMSVM for each positive example

Test time

Redundant Exemplars

Page 11: Visual Category as Collection of filters

Good Filters Bad Filters

Good / Bad Filters

Gradient orientation within a cell (active simultaneously)

Gradient orientation of neighboring cells (lines, curves)

Page 12: Visual Category as Collection of filters

Norm: consistent with high degree of alignment. Normalized Norm: Makes norm invariant to fi lter

dimension.

Cell Covariance: Diff erent orientation bins within a cell are highly structured. Gao et al. ECCV 2012

Cell Cross Covariance: Strong correlation between fi lter weights in nearby spatial locations.

Features for fi lter Ranking

Cell Covariance

Cell Cross Covariance

Decreasing Norm

Page 13: Visual Category as Collection of filters

– representation of fi lter Goal : model ranking score of by a linear functionTraining data : ,

where is number of training categories. where N is number of filters per category. is estimated quality, obtained by expensive method.

is ordered in descending value of - , for measures how much better is from

Slack rescaled hinge loss

Learning to Rank Filters

Page 14: Visual Category as Collection of filters

Selected parts should be individually good and complimentary.

First fi lter - fi lters selected so farSelect next fi lter using following

Greedy approximation for Diversity

Selected Filters

Not yet Selected

0.9

0.4

0.1

Added to selected set

Page 15: Visual Category as Collection of filters

LDA Acceleration

(w , λ)

N SVM filters(Candidate Generation)

n Selected Filters(SVM)

N LDA filters(Candidate Generation)

(w , λ)

n Selected Filters(LDA)

n Selected Filters(SVM)

SVM bootstrappi

ng

Poor performan

ce

Goodperforman

ce

Goodperforman

ce

Selection with LDA Acceleration

SVM bootstrapping

LDA

Our Selection Method

Page 16: Visual Category as Collection of filters

Experiments with Poselets

Test category

Filters used for training from remaining categories

800 poselet fi lters for each categoryGoal : given a category select 100 out of 800 fi ltersRanking taskDetection task

Page 17: Visual Category as Collection of filters

Performance of Ranker

Predicted ranking vs true ranking as per AP scores.

Norm(svm)

Σ – Norm(svm)

Gao et al. ECCV 2012

Rank(lda)

Rank(svm)< < <

Page 18: Visual Category as Collection of filters

1x

3x

8x

3x

3x

3x

8x

3x

3x

8x

2.4x

Detection Results

Σ – Norm (svm)

Random

Norm (svm)

10% Val

Norm (svm) + Div

Rank (svm)

Σ – Norm (svm) + Div

Oracle (expensive evaluation)

Rank (svm) + Div

Rank (lda) + Div

Rank (lda) + Div

2X Seeds

Increasin

g Perfo

rmance

Speed up w.r.t. OracleBy constructing a

poselet detector using selected fi lters

Order of magnitude Speed up.Improved performance than Oracle

Page 19: Visual Category as Collection of filters

Each category has 630 exemplars on average.Goal select 100 exemplars such that they reproduce

result for optimal set of 100 exemplars.Optimal set – weights of each exemplar in the final

scoring model. (Oracle)Frequency of exemplars

Experiments with exemplar SVMs

Frequent Exemplar

Rare Exemplar

Page 20: Visual Category as Collection of filters

We have presented an automatic mechanism for selecting diverse set of discriminative fi lters.

Order of magnitude improvement in training time.Our approach is applicable to any discriminative

architecture that uses a collection of fi lters. Insight into what makes a good fi lter for object

detection.Can be used as an attention mechanism during test

time Reduce number of convolutions / hashing lookups.

Bottom line: One can tell whether a fi lter is useful for a category without knowing what that category is, just by “looking” at the fi lter.