Fig. 3From: Consistent constraint-based video-level learning for action recognitionThe structure of the proposed networkBack to article page