Feature Hybrid Reconstruction Network for Few-Shot Action Recognition

The primary challenge in few-shot learning lies in acquiring generic characteristics from a limited number of labeled samples. Recent approaches in few-shot action recognition have achieved remarkable performance through the acquisition of discriminative features from each video during episodic training., as well as the implementation of diverse multi-feature fusion techniques. However, these methods often overlook the fact that the knowledge obtained from each episode may not be relevant in subsequent iterations, leading to potential failures in matching misaligned samples. In order to address these limitations, we propose a Feature Hybrid Reconstruction Network (FHRN) comprising a spatial reinforcement layer, a hybrid reconstruction module, and a prototypical matching strategy to effectively leverage spatial and temporal information. The FHRN adopts an agnostic approach to learn task-specific representations through the reconstruction process. Moreover, the set matching strategy enables precise identification of corresponding video frames across different videos. Experimental evaluations are conducted on various datasets including UCF10l, HMDB51, SSv2, and Kineticsl00. Our proposed FHRN method yields results that are comparable to, and in some cases surpass, those achieved by most state-of-the-art techniques.