Few-shot action recognition based on optical flow keyframes and hourglass convolution

Addressing the issues of potential loss of key action information and reduced data utilization efficiency due to current video sampling methods in the field of few-shot action recognition, as well as the inability of traditional feature extractors to fully exploit temporal information in videos, a new few-shot action recognition method based on optical flow keyframes (OFK) and hourglass convolution is proposed. Firstly, an OFK-based sampling strategy is designed to extract the most representative frames from videos, facilitating the capture of dynamic action features. Secondly, an efficient feature extractor (EFE) based on hourglass convolution is constructed, which can extract rich spatiotemporal features while maintaining computational efficiency. Finally, experiments were conducted on three video action datasets: HMDB51, UCF101, and Kinetics, achieving good classification accuracy and demonstrating the effectiveness of each module of the proposed method.