Convolutional Self-attention Guided Graph Neural Network for Few-Shot Action Recognition

The goal of few-shot action recognition is to recognize unseen action classes with only a few labeled videos. In this paper, we propose Convolutional Self-Attention Guided Graph Neural Network (CSA-GNN) for few-shot action recognition. First, for each video, we extract features of video frames sampled from the video and obtain a sequence of feature vectors. Then, a convolutional self-attention function is applied to the sequences to capture long-term temporal dependencies. Finally, a graph neural network is utilized to predict the distance between two sequences of feature vectors explicitly, which approximates the distance between the corresponding videos. By this means, we effectively learn the distance between the support and query videos without estimating their temporal alignment. The proposed method is evaluated on four action recognition datasets and achieves state-of-the-art results in the experiments.