Abstract: Few-shot action recognition aims to recognize novel action classes using just a few samples as knowledge. Most of the recent approaches learn to compare the similarity between videos. Recently, it has been observed that directly measuring this similarity is not ideal since the action instance cannot well aligned among videos. In this paper, we leverage the novel event boundary information to guide alignment learning in few-shot action recognition. First, a novel frame sampling strategy based on temporal boundaries is proposed to relieve the intra-class variance. Second, we propose a boundary selection module to locate the start & end time of action and further align videos to their duration. Ablation studies and visualizations demonstrate the effectiveness of the proposed methods. Extensive experiments on benchmark datasets show the potential of the proposed method in achieving state-of-the-art performance for few-shot action recognition.