A channel-wise spatial-temporal aggregation network for action recognition

A very challenging task for action recognition concerns how to effectively extract and utilize the temporal and spatial information of video (especially temporal information). To date, many researchers have proposed various spatial-temporal convolution structures. Despite their success, most models are limited in further performance especially on those datasets that are highly time-dependent due to their failure to identify the fusion relationship between the spatial and temporal features inside the convolution channel. In this paper, we proposed a lightweight and efficient spatial-temporal extractor, denoted as Channel-Wise Spatial-Temporal Aggregation block (CSTA block), which could be flexibly plugged in existing 2D CNNs (denoted by CSTANet). The CSTA Block utilizes two branches to model spatial-temporal information separately. In temporal branch, It is equipped with a Motion Attention Module (MA), which is used to enhance the motion regions in a given video. Then, we introduced a Spatial-Temporal Channel Attention (STCA) module, which could aggregate spatial-temporal features of each block channel-wisely in a self-adaptive and trainable way. The final experimental results demonstrate that the proposed CSTANet achieved the state-of-the-art results on EGTEA Gaze++ and Diving48 datasets, and obtained competitive results on Something-Something V1&V2 at the less computational cost.