Elastic temporal alignment for few‐shot action recognition

Few-shot action recognition aims to learn a classification model with good generalisation ability when trained only few labelled videos. However, it is difficult discriminative feature representations for videos in such setting. The Elastic Temporal Alignment (ETA) few-shot proposed. First, convolutional neural network employed extract of video frames sparsely sampled from In order obtain the similarity two videos, temporal alignment estimation function utilised estimate matching score between each pair through an elastic mechanism. analysis shows that we judge whether respective are matched, multiple adjacent should be considered, so as embody information. Thus, before feeding per-frame vectors into function, message passing leveraged propagate information features domain. method has been evaluated on four datasets, including Kinetics, Something-Something V2, HMDB51, and UCF101. experimental results verify effectiveness ETA show its superiority over state-of-the-art methods.