Noise Learning for Weakly Supervised Segment Classification in Video

This paper describes our solution for the 3rd YouTube-8M video understanding challenge. The challenge of this year is different from the previous challenge. Given a large scale video dataset with video-level labels and a small scale video dataset with segment-level labels, we are asked to recognize segments in videos this year. It can be regarded as a weakly supervised learning problem. To answer the challenge, we propose a solution consists of three different models, ie, segment-level classifier, self-attention mechanism, noise learning classifier. Among them, the noise learning classifier performs the best. By noise learning, it can reduce the noise of label and sample for training, and improve the performance. Moreover, we achieve the MAP of 0.78878 in the private leaderboard by model ensemble based on introduced models, ranking the 8th place on the challenge.