Multi-view Enhanced Few-shot Action Recognition for Embodied Visual Perception

With the advancement of embodied intelligence, robots are expected to understand human actions in dynamic, real-world environments with limited labeled data. Few-shot action recognition (FSAR) has thus emerged as a critical capability for enabling fast adaptation to new activities. While most current approaches focus on image-based tasks, extending few-shot learning to video is hindered by the complexity of spatiotemporal modeling and the lack of discriminative supervision. In this paper, we propose Multi-view Enhanced CLIP (MvE-CLIP), a novel framework for few-shot action recognition that leverages the strong visual-language priors of CLIP while enhancing its sensitivity to action-specific patterns. We introduce a multi-crop consistency strategy to generate multiple spatial views from video frames and explicitly model their similarities, encouraging the model to learn robust and consistent action representations. To evaluate the effectiveness of our approach, we conduct extensive experiments on two benchmark datasets, HMDB51 and UCF101, under standard few-shot settings. Ablation studies confirm the contribution of each component, and results show that MvE-CLIP outperforms state-of-the-art methods, particularly in challenging cases involving high visual similarity. Our work demonstrates the potential of combining pre-trained vision-language models with multi-view consistency for efficient and generalizable action understanding in data-scarce scenarios.