Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.
CVPR 2009, Miami, Florida Subhransu Maji and Jitendra Malik University of California at Berkeley, Berkeley, CA-94720 Objec...
Overview <ul><li>Overview of probabilistic Hough transform </li></ul><ul><li>Learning framework </li></ul><ul><li>Experime...
Our Approach: Hough Transform <ul><li>Popular for detecting parameterized shapes </li></ul><ul><ul><li>Hough’59, Duda&Hart...
Generalized to object detection  Learning <ul><li>Learn appearance codebook </li></ul><ul><ul><li>Cluster over interest po...
Detection Pipeline B. Leibe, A. Leonardis, and B. Schiele.  Combined object categorization and segmentation with an implic...
Probabilistic Hough Transform <ul><li>C – Codebook </li></ul><ul><li>f – features, l - locations </li></ul>Position Poster...
Learning Feature Weights <ul><li>Given  :  </li></ul><ul><ul><li>Appearance Codebook, C  </li></ul></ul><ul><ul><li>Poster...
<ul><li>Naïve Bayes weights: </li></ul><ul><li>Encourages relatively rare parts </li></ul><ul><li>However rare parts may n...
<ul><li>Location invariance assumption </li></ul><ul><li>Overall score is linear given the matched codebook entries </li><...
Max-Margin Training <ul><li>Training:  </li></ul><ul><li>Construct dictionary  </li></ul><ul><li>Record codeword distribut...
Experiment Datasets ETHZ Shape Dataset ( Ferrari et al., ECCV 2006)  255 images, over 5 classes (Apple logo, Bottle, Giraf...
Experimental Results <ul><li>Hough transform details </li></ul><ul><ul><li>Interest points : Geometric Blur descriptors at...
Learned Weights (ETHZ shape) Max-Margin Important Parts Naïve Bayes blue (low)  ,  dark red (high) Influenced by clutter (...
Learned Weights (UIUC cars) blue (low)  ,  dark red (high) Naïve Bayes Max-Margin Important Parts
Learned Weights (INRIA horses) blue (low)  ,  dark red (high) Naïve Bayes Max-Margin Important Parts
Detection Results (ETHZ dataset) Recall @ 1.0 False Positives Per Window
Detection Results (INRIA Horses) Our Work
Detection Results (UIUC Cars) INRIA horses Our Work
Hough Voting + Verification Classifier Recall @ 0.3 False Positives Per Image  ETHZ Shape Dataset  IKSVM was run on top 30...
Hough Voting + Verification Classifier IKSVM was run on top 30 windows + local search Our Work
Hough Voting + Verification Classifier UIUC Single Scale Car Dataset IKSVM was run on top 10 windows + local search 1.7% i...
Summary <ul><li>Hough transform based detectors offer good detection performance and speed.  </li></ul><ul><li>To get bett...
<ul><li>Work partially supported by: </li></ul><ul><li>ARO MURI W911NF-06-1-0076 and ONR MURI N00014-06-1-0734 </li></ul><...
Backup Slide : Toy Example Rare but poor localization Rare and good localization
Upcoming SlideShare
Loading in …5
×

CVPR2009: Object Detection Using a Max-Margin Hough Transform

2,083 views

Published on

Published in: Education
  • Be the first to comment

CVPR2009: Object Detection Using a Max-Margin Hough Transform

  1. 1. CVPR 2009, Miami, Florida Subhransu Maji and Jitendra Malik University of California at Berkeley, Berkeley, CA-94720 Object Detection Using a Max-Margin Hough Transform
  2. 2. Overview <ul><li>Overview of probabilistic Hough transform </li></ul><ul><li>Learning framework </li></ul><ul><li>Experiments </li></ul><ul><li>Summary </li></ul>
  3. 3. Our Approach: Hough Transform <ul><li>Popular for detecting parameterized shapes </li></ul><ul><ul><li>Hough’59, Duda&Hart’72, Ballard’81,… </li></ul></ul><ul><li>Local parts vote for object pose </li></ul><ul><li>Complexity : # parts * # votes </li></ul><ul><ul><li>Can be significantly lower than brute force search over pose (for example sliding window detectors) </li></ul></ul>
  4. 4. Generalized to object detection Learning <ul><li>Learn appearance codebook </li></ul><ul><ul><li>Cluster over interest points on </li></ul></ul><ul><ul><li>training images </li></ul></ul><ul><li>Use Hough space voting to find objects </li></ul><ul><ul><li>Lowe’99, Leibe et.al.’04,’08, Opelt&Pinz’08 </li></ul></ul><ul><li>Implicit Shape Model </li></ul><ul><ul><li>Leibe et.al.’04,’08 </li></ul></ul><ul><li>Learn spatial distributions </li></ul><ul><ul><li>Match codebook to training images </li></ul></ul><ul><ul><li>Record matching positions on object </li></ul></ul><ul><ul><li>Centroid is given </li></ul></ul>Spatial occurrence distributions x y s x y s x y s x y s
  5. 5. Detection Pipeline B. Leibe, A. Leonardis, and B. Schiele. Combined object categorization and segmentation with an implicit shape model ‘ 2004 Probabilistic Voting Interest Points eg. SIFT,GB, Local Patches Matched Codebook Entries KD Tree
  6. 6. Probabilistic Hough Transform <ul><li>C – Codebook </li></ul><ul><li>f – features, l - locations </li></ul>Position Posterior Codeword Match Codeword likelihood Detection Score Codeword likelihood
  7. 7. Learning Feature Weights <ul><li>Given : </li></ul><ul><ul><li>Appearance Codebook, C </li></ul></ul><ul><ul><li>Posterior distribution of object center for each codeword P(x|…) </li></ul></ul><ul><li>To Do : </li></ul><ul><ul><li>Learn codebook weights such that the Hough transform detector works well (i.e. better detection rates) </li></ul></ul><ul><li>Contributions : </li></ul><ul><ul><li>Show that these weights can be learned optimally using a max-margin framework. </li></ul></ul><ul><ul><li>Demonstrate that this leads to improved accuracy on various datasets </li></ul></ul>
  8. 8. <ul><li>Naïve Bayes weights: </li></ul><ul><li>Encourages relatively rare parts </li></ul><ul><li>However rare parts may not be good predictors of the object location </li></ul><ul><li>Need to jointly consider both priors and distribution of location centers. </li></ul>Learning Feature Weights : First Try
  9. 9. <ul><li>Location invariance assumption </li></ul><ul><li>Overall score is linear given the matched codebook entries </li></ul>Learning Feature Weights : Second Try Position Posterior Codeword Match Codeword likelihood Activations Feature weights
  10. 10. Max-Margin Training <ul><li>Training: </li></ul><ul><li>Construct dictionary </li></ul><ul><li>Record codeword distributions on training examples </li></ul><ul><li>Compute “a” vectors on positive and negative training examples </li></ul><ul><li>Learn codebook weights using by max-margin training </li></ul>Standard ISM model (Leibe et.al.’04) Our Contribution class label {+1,-1} activations non negative
  11. 11. Experiment Datasets ETHZ Shape Dataset ( Ferrari et al., ECCV 2006) 255 images, over 5 classes (Apple logo, Bottle, Giraffe, Mug, Swan) UIUC Single Scale Cars Dataset ( Agarwal & Roth, ECCV 2002) 1050 training, 170 test images INRIA Horse Dataset ( Jurie & Ferrari) 170 positive + 170 negative images (50 + 50 for training)
  12. 12. Experimental Results <ul><li>Hough transform details </li></ul><ul><ul><li>Interest points : Geometric Blur descriptors at sparse sample of edges (Berg&Malik’01) </li></ul></ul><ul><ul><li>Codebook constructed using k -means </li></ul></ul><ul><ul><li>Voting over position and aspect ratio </li></ul></ul><ul><ul><li>Search over scales </li></ul></ul><ul><li>Correct detections (PASCAL criterion) </li></ul>
  13. 13. Learned Weights (ETHZ shape) Max-Margin Important Parts Naïve Bayes blue (low) , dark red (high) Influenced by clutter (rare structures)
  14. 14. Learned Weights (UIUC cars) blue (low) , dark red (high) Naïve Bayes Max-Margin Important Parts
  15. 15. Learned Weights (INRIA horses) blue (low) , dark red (high) Naïve Bayes Max-Margin Important Parts
  16. 16. Detection Results (ETHZ dataset) Recall @ 1.0 False Positives Per Window
  17. 17. Detection Results (INRIA Horses) Our Work
  18. 18. Detection Results (UIUC Cars) INRIA horses Our Work
  19. 19. Hough Voting + Verification Classifier Recall @ 0.3 False Positives Per Image ETHZ Shape Dataset IKSVM was run on top 30 windows + local search KAS – Ferrari et.al., PAMI’08 TPS-RPM – Ferrari et.al., CVPR’07 better fitting bounding box Implicit sampling over aspect-ratio
  20. 20. Hough Voting + Verification Classifier IKSVM was run on top 30 windows + local search Our Work
  21. 21. Hough Voting + Verification Classifier UIUC Single Scale Car Dataset IKSVM was run on top 10 windows + local search 1.7% improvement
  22. 22. Summary <ul><li>Hough transform based detectors offer good detection performance and speed. </li></ul><ul><li>To get better performance one may learn </li></ul><ul><ul><li>Discriminative dictionaries (two talks ago, Gall et.al.’09) </li></ul></ul><ul><ul><li>Weights on codewords (our work) </li></ul></ul><ul><li>Our approach directly optimizes detection performance using a max-margin formulation </li></ul><ul><li>Any weak predictor of object center can be used is this framework </li></ul><ul><ul><li>Eg. Regions (one talk ago, Gu et.al. CVPR’09) </li></ul></ul>
  23. 23. <ul><li>Work partially supported by: </li></ul><ul><li>ARO MURI W911NF-06-1-0076 and ONR MURI N00014-06-1-0734 </li></ul><ul><li>Computer Vision Group @ UC Berkeley </li></ul>Acknowledgements Thank You Questions?
  24. 24. Backup Slide : Toy Example Rare but poor localization Rare and good localization

×