Feature Sequence Representation Via Slow Feature Analysis For Action Classification

The recent advances in extracting motion descriptors, such as BoW and CNN features, enable us to effectively convert a video into a sequence of frame-based feature vectors. For improving the action classification performance, in this paper, we propose an efficient method to represent the feature sequence by exploiting the temporal patterns via slow feature analysis (SFA). The ordinary SFA suffers from small sample size (SSS) problem found in action video clips and thus we propose PCA-SFA to cope with the SSS problem by incorporating the information of PCA subspaces into SFA. The proposed method leverages the PCA-SFA projection vector to describe the sequence of even fewer frames by a fixed-dimensional video descriptor, capturing the essential temporal dynamics which is a slowly varying pattern embedded in the quickly varying input signals. The computational cost to produce the video descriptor is negligible compared to the feature extraction process such as BoW and CNN since the PCA-SFA is computed in a computationally efficient manner. In the experiments on action classification using various datasets, the proposed method exhibits favorable performance being competitive to the other methods.