Multiple Feature Learning for Action Classification Goal: Classify actions in images Approach SIFT, HOG, LBP, and color histogram features Use dense samping, LLC, and SPM to get histogram intersection kernel for each feature Compared stacking, feature selection, and randomization techniques to kernel learning Combine kernels using multiple kernel learning: Riding horse K1 K2 Kfinal C-SVM Km Multiple Kernel Learning From Wang et al., 2010 Ben Poole (Computer Science Department, Stanford University) 1
Multiple Feature Learning for Action Classification Combining features improves performance by 10% Stacking feature vectors yields similar performance to MKL Randomly selected max features (WTA-Hash) also achieve similar performance for a smaller computational cost Ben Poole (Computer Science Department, Stanford University) 2