Video Action Recognition

Video action recognition studies how to identify activities from sequences of frames, where timing and interactions can distinguish actions that look similar in a single image. Common approaches include 3D convolutional networks and video transformers, RGB and optical-flow streams, pose or skeleton models, and pretrained vision-language models adapted through prompts or limited fine-tuning. Research also addresses few-shot and open-vocabulary recognition, temporal localization, and transfer across domains such as egocentric, medical, sports, and surveillance video. A central challenge is improving temporal understanding and generalization under viewpoint changes, occlusion, scarce labels, and background shortcuts while keeping models efficient and reliable in real-world settings.