Few-Shot Video Recognition via Hierarchical Metric Learning

To address the issues of existing few-shot action recognition methods that rely on single-prototype supervision, struggle to utilize global spatial information across frames in videos, and have limited generalization ability of prototypes, researchers proposed a new method called Hierarchical Metric Learning for Few-Shot Action Recognition (HML-FSAR). This method first captures global spatial representations across frames through a spatial enhancement module, and then constructs a complete feature processing pipeline by combining a time-based multi-head attention mechanism, heterogeneous alignment, spatio-temporal feature fusion, and dictionary learning modules. Subsequently, a hierarchical metric learning strategy consisting of central metric, alignment metric, contrast metric, dictionary metric, and prototype metric is embedded into it. Progressive multi-stage complementary constraints are applied from the frame-level representation to the final prototype, jointly optimizing feature compactness, heterogeneous spatio-temporal alignment, inter-class discrimination, and noise resistance. This method has been validated on five widely used few-shot action recognition datasets, and experimental results demonstrate its effectiveness.