An HOG-LBP Human Detector with Partial Occlusion Handling<br />  Xiaoyu Wang*, Tony X. Han*, and Shuicheng Yan†<br />*ECE ...
Human detection, or more generally, object detection, has wide applications<br />Currently, Sliding Window Classifiers (SW...
An HOG-LBP Human Detector with Partial Occlusion Handling<br />3<br />Outline<br />The proposed HOG-LBP feature<br />Parti...
HOG and LBP feature<br />Traditional HOG Feature -N. Dalal and B. Triggs. Histograms of oriented gradients for human detec...
Cell-structured LBP designed especially for human detection<br />Holistic LBP histogram for each sliding window achieves p...
The performance of cell-structured LBP <br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br /...
Classifier:Linear SVM</li></ul>         HOG  <br />
HOG-LBP feature<br />Why simple concatenation helps?<br />Disadvantage of HOG:<br />Focusing on edge, ignoring flat area<b...
The performance of HOG-LBP feature <br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />8<b...
HOG-LBP feature for general object detection<br />The proposed HOG-LBP feature works pretty well for general object detect...
Two key questions<br />Does the partial occlusion occur in the current scanning window?<br />If partial occlusion occurs, ...
Convert holistic classifier to local-classifier ensemble<br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlus...
Distribute the constant bias to local classifiers<br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Han...
Segmenting the local classifiers for occlusion inference<br />The over all occlusion reasoning/handling framework.<br />9/...
The detection performance with occlusion handling<br />Samples of corrected miss detection<br />9/28/2009<br />An HOG-LBP ...
There are very few occluded pedestrians in INRIA dataset.
28 images with occlusion are missed  by HOG-LBP detector when FPPW=10-6
The occlusion handling pickup 10 of them.</li></li></ul><li>Adding occlusions to  INRIA dataset <br />9/28/2009<br />An HO...
Evaluation using False Positive Per scanning Imange (FPPI)<br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occl...
9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />17<br />Pascal 2009 Grand Challenge<br />pre...
Sample results in Geoint 2009<br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />18<br />
Upcoming SlideShare
Loading in...5
×

Human detection iccv09

2,933

Published on

Presentation for the paper "An HOG -LBP" human detector with partial occlusion handling

Published in: Career
0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total Views
2,933
On Slideshare
0
From Embeds
0
Number of Embeds
1
Actions
Shares
0
Downloads
107
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide

Human detection iccv09

  1. 1. An HOG-LBP Human Detector with Partial Occlusion Handling<br /> Xiaoyu Wang*, Tony X. Han*, and Shuicheng Yan†<br />*ECE Department University of Missouri, Columbia, MO, USA<br />† ECE Department National University of Singapore, Singapore<br />
  2. 2. Human detection, or more generally, object detection, has wide applications<br />Currently, Sliding Window Classifiers (SWC) achieves the best performance in object detection<br />“Sliding window classifier predominant”(Everinghamet al. The PASCAL Visual Object Classes Challenge workshop 2008, 2009)<br />-“HOG tends to outperform other methods surveyed,”(Dollar et al. “Pedestrian Detection: A Benchmark”, CVPR2009)<br />But still, lots of things need to be improved for SWCs <br />More robust features are always desirable<br />Compared with part-based detector, sliding window approach handles occlusion poorly<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />2<br />Introduction<br />9/28/2009<br />Binary Classifier<br />Pos: patch with a human<br />Neg: patch with no human<br />
  3. 3. An HOG-LBP Human Detector with Partial Occlusion Handling<br />3<br />Outline<br />The proposed HOG-LBP feature<br />Partial occlusion handling<br />Results and performance evaluation<br />The speed: making it real-time!<br />Conclusion and real-time demo<br />9/28/2009<br />
  4. 4. HOG and LBP feature<br />Traditional HOG Feature -N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In CVPR 2005, vol. 1, pp. 886–893, 2005.<br />Traditional Local Binary Pattern (LBP) feature<br />LBP operator is an exceptional texture descriptors<br />LBP has achieved good results in face recognitionT. Ahonen, et al. Face description with local binary patterns: Application to face recognition. IEEE PAMI, 28(12):2037–2041, 2006.<br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />4<br />
  5. 5. Cell-structured LBP designed especially for human detection<br />Holistic LBP histogram for each sliding window achieves poor results.<br />Inspired by the success of the HOG, LBP histograms are constructed for each cell with the size 16by16<br />In contrast to HOG, no block structure is needed if we use L1 normalization.<br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />5<br />…<br />
  6. 6. The performance of cell-structured LBP <br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />6<br />Missing rate vs. False Positive Per scanning Window (FPPW)<br />Results on INRIA dataset<br /><ul><li>Feature:Cell-structured LBP
  7. 7. Classifier:Linear SVM</li></ul> HOG <br />
  8. 8. HOG-LBP feature<br />Why simple concatenation helps?<br />Disadvantage of HOG:<br />Focusing on edge, ignoring flat area<br />Can not deal with noisy edge region <br />Advantage of Cell-LBP: <br />Treat all the patterns equally <br />Filter out noisy patterns using the concept of “uniform patterns ”, i.e. vote all strings with more than k 0-1 transition into same bin.<br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />7<br />
  9. 9. The performance of HOG-LBP feature <br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />8<br />Missing rate vs. FPPW<br />[1] N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” in CVPR, 2005.<br />[2] O. Tuzel, F. Porikli, and P. Meer, “Human detection via classification on Riemannian manifolds,” in CVPR 2007.<br />[3] S. Maji, A. Berg, and J. Malik, “Classification using intersection kernel support<br />vector machines is efficient,” in CVPR 2008.<br />[4] HOG-LBP without occlusion handling<br />
  10. 10. HOG-LBP feature for general object detection<br />The proposed HOG-LBP feature works pretty well for general object detection.<br />We attended the Pascal 2009 grand challenge in object detection. Among 20 categories, using the HOG-LBP as feature, our team (Mizzou) got:<br />Number 1 in two categories: chair, potted plant<br />Number 2 in four categories: bottle, car, person, horse<br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />9<br />
  11. 11. Two key questions<br />Does the partial occlusion occur in the current scanning window?<br />If partial occlusion occurs, where?<br />An interesting phenomenon<br />Partial occlusion handling<br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />10<br />Negative Positive<br /><hP, hU><br /><hN, hL><br />Negative Positive<br />
  12. 12. Convert holistic classifier to local-classifier ensemble<br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />11<br />?<br />
  13. 13. Distribute the constant bias to local classifiers<br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />12<br />positive training samples<br />negative training samples<br />the feature of the ith blocks of<br />the feature of the ith blocks of<br />This approach of distributing the constant bias keeps the relative bias ratio across the whole training dataset.<br />
  14. 14. Segmenting the local classifiers for occlusion inference<br />The over all occlusion reasoning/handling framework.<br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />13<br />
  15. 15. The detection performance with occlusion handling<br />Samples of corrected miss detection<br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />14<br /><ul><li>The detection rate improvement is less than 1% for INRIA Dataset.
  16. 16. There are very few occluded pedestrians in INRIA dataset.
  17. 17. 28 images with occlusion are missed by HOG-LBP detector when FPPW=10-6
  18. 18. The occlusion handling pickup 10 of them.</li></li></ul><li>Adding occlusions to INRIA dataset <br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />15<br />
  19. 19. Evaluation using False Positive Per scanning Imange (FPPI)<br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />16<br />[1] P. Sabzmeydani and G. Mori. Detecting pedestrians by learning shapelet features.<br />In CVPR 2007.<br />[2] P. Dollar, Z. Tu, H. Tao, and S. Belongie. Feature mining for image classification. In CVPR 2007<br />[3] S. Maji, A. Berg, and J. Malik, “Classification using intersection kernel support vector machines is efficient,” in CVPR 2008.<br />[4] N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” in CVPR, 2005.<br />[5] P. Felzenszwalb, D. McAllester, and D. Ramanan. A discriminatively trained,<br />multiscale, deformable part model. In CVPR, 2008.<br />[6] C.Wojek and B. Schiele. A performance evaluation of single and multi-feature people detection. DAGM 2008. <br />[7], [8] HOG-LBP w/o occlusion handling<br />
  20. 20. 9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />17<br />Pascal 2009 Grand Challenge<br />precision<br />Average Precision:<br />UoCTTI: 41.5<br />U of Missouri: 37.0<br />Oxford_MKL: 21.6<br />recall<br />
  21. 21. Sample results in Geoint 2009<br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />18<br />
  22. 22. Evaluation Issue<br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />19<br />Many factors affect FFPI:Like nonmaximum suppression, bandwidth of meanshift, local thresholding/filtering before merging.<br />Therefore:<br />Using FPPW for sliding window classifier to select feature and classification scheme. <br />WARNING: avoid encoding the class label implicitly<br />Using FPPI to evaluate the over all performance of the detector, can be used as a protocol to compare all kinds of detectors<br />
  23. 23. Speed Issue: do trilinear Interpolation as convolution<br />9/28/2009<br />An HOG-LBP Human Detector with Partial Occlusion Handling<br />20<br />Linear interpolation<br />Trilinear interpolation<br />Trilinear interpolation can now be integrated into integral histogram, and improve the detection by 3%-4%, at FPPW=10-4.<br />Adjacent histograms cover independent data after convolution. SPMD, this is very important if you want to use GPU! Memory bandwidth is more precious than GPU cycles.<br />
  24. 24. An HOG-LBP Human Detector with Partial Occlusion Handling<br />21<br />Conclusion and Demo<br />The HOG-LBP feature achieves the state of the art detection.<br />Segmentation on local classifications inside sliding window helps to infer occlusion.<br />Implementing trilinear interpolation as a 2D convolution makes it an addable component of integral histogram.<br />Demo <br />Does it work? Press keyboard and pray......<br />We may still have long way to go<br />9/28/2009<br />
  1. ¿Le ha llamado la atención una diapositiva en particular?

    Recortar diapositivas es una manera útil de recopilar información importante para consultarla más tarde.

×