Action Recognition Using Accelerated Local Descriptors and Temporal Variation

Our system performs late fusion of several features whose weights have been optimized on UCF50 dataset. The fusion is done over the following features: 1) Our newly developed fast local descriptors for HoG and HoF, both grey-scale and RGB. In RGB-HoG/HoF we compute the dense HoG and HoF descriptors for all color channels and concatenate them. To obtain a single vector per video, we use the Fisher kernel with 256 centers using power normalization and L2 norm. 2) Global HoF features extracted for each video frame. The temporal variation has been modeled by using the Fisher Kernel representation, shown to be beneficial in [1]. Finally, we apply the late fusion over the outputs of these two methods on the UCF101 dataset.