Temporal action localization is a challenging task in video understanding. Although great progress has been made in temporal action localization, the most advanced methods still have the problem of sharp performance degradation when an action proposal generated. Most methods use sliding windows method or simply group frames according to frame-level scores. These methods are not enough to provide accurate action boundary and maintain reasonable temporal structure. In order to solve these problems, we propose a novel proposal optimization network to generate start score, end score, action score and regression score, and then remove the redundancy by NMS algorithm. In the proposed method, we introduce a metric loss function to maintain the temporal structure of action proposal in the training process. To verify the effectiveness of the proposed method, we have made comparative experiments on ActivityNet-1.3 dataset respectively, and the proposed method has surpassed some of the state-of-the-art methods on the dataset.
展开▼
