470 likes | 604 Views
“Secret” of Object Detection. Zheng Wu (Summer intern in MSRNE) Sep. 3, 2010 Joint work with Ce Liu (MSRNE) William T. Freeman (MIT) Adam Kalai (MSRNE) Jianxiong Xiao (MIT). person. motorbike. Outline. Introduction Sliding-Window-based Object Detection Window Generation
E N D
“Secret” of Object Detection Zheng Wu (Summer intern in MSRNE) Sep. 3, 2010 Joint work with Ce Liu (MSRNE) William T. Freeman (MIT) Adam Kalai (MSRNE) Jianxiong Xiao (MIT)
person motorbike
Outline Introduction Sliding-Window-based Object Detection Window Generation Feature Extraction Cascade Classifier PASCAL Challenge Conclusion
Related Topic • Image Matching • Image Classification • Object Detection
Object Detection • Single Pattern • Multiple Patterns Viola & Jones Face Detector Felzenszwalb’s Part-based Detector Dalal & Triggs Pedestrian Detector
Outline Introduction Sliding-Window-based Object Detection Window Generation Feature Extraction Cascade Classifier PASCAL Challenge Conclusion
Sliding Windows 7 • Search over: • Location • Scale • Aspect Ratio • Millions of windows!
Sliding Windows • Subsample on grid • - have to set “optimal” step size manually • Fix Aspect Ratio - assume single pattern detection • Fix Scale - assume object’s resolution does not change much between training and test sets. • Search with branch-and-bound method • - have to use special scoring function
Sliding Windows • We propose sliding windows from segmentation Region Segmentation [Felzenszwalb & Huttenlocher, IJCV03] SuperpixelSegmentation [Levinshtein et al, PAMI09] < 100,000 sliding windows / image on PASCAL Dataset
Outline • Introduction • Sliding-Window-based Object Detection • Window Generation • Feature Extraction • Cascade Classifier • PASCAL Challenge • Conclusion
Generic Feature • “Objectness” features [Alexe et al, CVPR10]
Generic Feature • Each type of generic feature is weak, but combination is stronger • Low dimensional feature (=8) • Not suitable for objects with “concave” shape, i.e. table, chair
Class-specific Feature • Histogram of Orientated Gradients
Class-specific Feature • Dense grid (>10*10) (secret 1) • Overlapping cells • Histogram bin size • High dimensional feature (>1000) • - redundant oroverfitting? • Normalization • No spatial relationship maintained
Outline • Introduction • Sliding-Window-based Object Detection • Window Generation • Feature Extraction • Cascade Classifier • PASCAL Challenge • Conclusion
Cascade Classifier • Same type of classifier with different features • Viola & Jones Face Detector, IJCV01 • Different types of classifier with same features • Harzallah et al, ICCV09 (INRIA) • Vedaldi et al, ICCV09 (Oxford)
Cascade Classifier • Training SVM is slow… • to train 20,000 examples with 4000 dimensions: • >15min for Linear SVM • >3 hours for Nonlinear SVM • Training SVM requires a lot of memory… • design matrix: 20,000*20,000 matrix • Training with Imbalance data • a few hundreds of positive examples • billions of negative examples
Boosted SVM • 42327 examples, half for training, half for testing • Training error is 0.05 for all boosted classifiers
Negative Training Set Training Sample Pool False Positives Random Samples … … Secret 2 SVM ver. 2 SVM ver. 1
Outline • Introduction • Sliding-Window-based Object Detection • Window Generation • Feature Extraction • Cascade Classifier • PASCAL Challenge • Conclusion
Conclusion • Proposing sliding windows without fixing scale or aspect ratio is possible. • Simple feature (saliency, contrast, etc) is only useful for certain object classes. • Histogram-based feature is not sufficient to represent real world object, no matter what learning algorithm is used. • Boosting is helpful to speed up SVM-training and reduce the memory usage. • Digging out “hard” negative examples. • Throwing away “hard” positive examples.
Future Work • It is time to go beyond the histogram-of-X feature • - not every pixel within bounding box is informative • - the appearance of object’s part is more robust • Evolve the classifier • - even PASCAL dataset is too small • - the right decision boundary is still far away… • - Active learning? Online learning?