Abstract: Few-shot video action recognition aims to construct efficient learning models using limited training samples,thereby reducing the dependence of traditional action recognition on large-scale and finely annotated datasets.At present,most few-shot learning models classify videos based on their similarity.However,due to the different spatiotemporal distributions of action instances,there is a temporal and action evolution mismatch between the query video and the supporting video,which affects the recognition performance of the model.To address this issue,a two-stage spatiotemporal alignment network TSAN is proposed to improve the alignment accuracy of video data,thereby enhancing the accuracy of few-shot video action recognition.This network adopts the basic architecture of meta learning.In the first stage,the action time alignment module ATAM is used to construct video frame pairs in tuple mode,which subdivides video actions into sub action sequences and combines them with temporal information in video data to improve the efficiency of few-shot learning.In the second stage,the action evolution alignment module AEAM,along with its time synchronization submodule TSM and spatial coordination submodule SCM,are used to calibrate the query features to match the spatiotemporal action evolution of the support set,thereby improving the accuracy of few-shot video action recognition.The experimental results on the HMDB51,UCF101,SSV2100,and Kinetics100 datasets show that the TSAN network has higher recognition accuracy compared to existing few-shot video action recognition methods.

Key words: Action recognition, Video classification, Spatio-temporal alignment, Few-shot learning, Meta-learning