UPL-Net: Uncertainty-aware prompt learning network for semi-supervised action recognition

This paper focuses on understanding human behavior in videos by reframing the traditional video classification task as a transfer learning problem centered on visual concepts. Unlike existing action recognition approaches that rely solely on single-modal representations and video classifiers, our method leverages an uncertainty- aware prompt learning network (UPL-Net). This network is designed to extract spatiotemporal features that are pertinent to action-related concepts in videos while ensuring that the visual concepts derived from images are preserved. Furthermore, we introduce an uncertainty-guided semi-supervised learning strategy that harnesses unlabeled videos to enhance the model's generalizability. Extensive experiments conducted on benchmark datasets, namely UCF and HMDB, demonstrate the superiority of our approach over state-of-the-art semi- supervised action recognition methods. Notably, under a 1% labeling rate on the UCF dataset, our method achieves a significant improvement of 12.8%, underscoring its effectiveness in leveraging limited labeled data and abundant unlabeled videos for improved performance.