Spatio-Temporal Modeling for Action Recognition in Videos

Technological innovation in the field of video action recognition drives the development of video-based real-world applications. This PhD thesis provides a new set of machine learning algorithms for processing videos efficiently, leading to outstanding results in human action recognition in videos. First of all, two video representation extraction methods, Temporal Squeezed Pooling (TSP) and Pixel-Wise Temporal Projection (PWTP), are proposed in order to enhance the discriminative video feature learning abilities of Deep Neural Networks (DNNs). TSP enables spatio-temporal modeling by temporally aggregating the information from long video frame sequences. PWTP is an improved version TSP, which filters out static appearance while performing information aggregation. Secondly, we discuss how to address the long-term dependency modeling problem of video DNNs. To this end, we develop two spatio-temporal attention mechanisms, Region-based Non-local (RNL) and Convolution Pyramid Attention (CPA). We devise an attention chain by connecting the RNL or CPA module to the Squeeze-Excitation (SE) operation. We demonstrate how the attention mechanisms can be embedded into deep networks to alleviate the optimization difficulty. Finally, we are focused on tackling the problem of heavy computational cost in video models. To this end, we introduce the concept of busy-quiet video disentangling for exceedingly fast video modeling. We propose the Motion Band-Pass Module (MBPM) embedded into the Busy-Quiet Net (BQN) architecture to reduce videos’ information redundancy in the spatial and temporal dimensions. The BQN architecture is extremely lightweight while still performing better than other heavier models. Extensive experiments for all the proposed methods are provided on multiple video benchmarks, including UCF101, HMDB51, Kinetics400, Something-Something V1.