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 ID: 697076
Download Presentation The PPT/PDF document "Knowing a Good HOG Filter" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Knowing a Good HOG Filter When You See It:Efficient Selection of Filters for Detection Ejaz Ahmed1, Gregory Shakhnarovich2, and Subhransu Maji31 University of Maryland, College Park2 Toyota Technological Institute at Chicago3 University of Massachusetts, AmherstSlide2
Visual Category as Collection of filtersPoseletsMid Level Discriminative PatchesExemplar SVMsSlide3
Candidate Generation
Candidate Generation
Pool of Filters
Filters are generated using positives and negatives examples.
Generation of a large pool of filters.
Positives
Negatives
Filter
(
SVM Classifier)Slide4
Two sources of inefficiencyCandidate SelectionCandidateSelection
Selected Filters (n)(n << N)
Pool of Filters (N)
Impractical to use all generated filters.
Redundancy
Noise
good
badSlide5
Candidate Selection Cont…Expensive EvaluationSelected Filters (n)
(n << N)
Run as detector
Pool of Filters (N)
BottleneckSlide6
What we ProposeExpensive EvaluationSelected Filters (n)(n << N)
Run as detector
Pool 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 filtersSlide7
Category Independent Model
Images +/-
Pool (Candidate Filters)
N >> n
Selected Filters (n)
fast
slow
(w ,
λ
)
Test Category
fast
N candidates
n selected
fast
C
an rank filters as accurately as a direct evaluation on thousands of examples.Slide8
PoseletsPoselets are semantically aligned discriminative patterns that capture parts of object.
Patches are often far visually, but they are close semanticallySlide9
Poselet ArchitectureCandidate Selection :
Candidate Generation :Slide10
Save significantly in training time if we can quickly select small set of relevant exemplars.ESVMSVM for each positive example
Test time
Redundant ExemplarsSlide11
Good FiltersBad FiltersGood / Bad Filters
G
radient orientation within a cell (active simultaneously)
G
radient orientation of neighboring cells (lines, curves)Slide12
Norm: consistent with high degree of alignment. Normalized Norm: Makes norm invariant to filter dimension. Cell Covariance: Different orientation bins within a cell are highly structured. Gao et al. ECCV 2012Cell Cross Covariance: Strong correlation between filter weights in nearby spatial locations.Features for filter Ranking
Cell Covariance
Cell Cross Covariance
Decreasing NormSlide13
– representation of filter 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 FiltersSlide14
Selected parts should be individually good and complimentary.First filter - filters selected so farSelect next filter using following Greedy approximation for Diversity
Selected Filters
Not yet Selected
0.9
0.4
0.1
Added to selected setSlide15
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 bootstrapping
Poor performance
Good
performance
Good
performance
Selection with
LDA Acceleration
SVM bootstrapping
LDA
Our Selection MethodSlide16
Experiments with Poselets
Test category
Filters used for training from remaining categories
800
poselet
filters for each category
Goal : given a category select 100 out of 800 filters
Ranking task
Detection taskSlide17
Performance of RankerPredicted ranking vs true ranking as per AP scores.Norm(svm)Σ – Norm(svm)Gao et al. ECCV 2012Rank(lda)Rank(svm)<<<Slide18
1x3x8x3x3x3x8x3x
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) +
DivRank (
lda) + Div2X Seeds
Increasing PerformanceSpeed up w.r.t. OracleBy constructing a poselet detector using selected filters
Order of magnitude Speed up.
Improved performance than OracleSlide19
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 exemplarsExperiments with exemplar SVMsFrequent ExemplarRare ExemplarSlide20
We have presented an automatic mechanism for selecting diverse set of discriminative filters.Order of magnitude improvement in training time.Our approach is applicable to any discriminative architecture that uses a collection of filters.Insight into what makes a good filter for object detection.Can be used as an attention mechanism during test timeReduce number of convolutions / hashing lookups.Conclusion
Bottom line: One can tell whether a filter is useful for a category without knowing what that category is, just by “looking” at the filter.