Intra-Camera Supervised Person Re-Identification

Abstract Existing person re-identification (re-id) methods mostly exploit a large set of cross-camera identity labelled training data. This requires tedious data collection and annotation process, leading to poor scalability in practical re-id applications. On the other hand unsupervised do not need label information, but they usually suffer from much inferior insufficient model performance. To overcome these fundamental limitations, we propose novel paradigm based on an idea independent per-camera annotation. eliminates most time-consuming inter-camera labelling significantly reducing amount human efforts. Consequently, it gives rise more scalable feasible setting, which call Intra-Camera Supervised (ICS) re-id, for formulate Multi-tAsk mulTi-labEl (MATE) deep learning method. Specifically, MATE is designed self-discovering correspondence multi-task inference framework. Extensive experiments demonstrate cost-effectiveness superiority our method over alternative approaches three datasets. For example, yields 88.7% rank-1 score Market-1501 proposed ICS outperforming models closely approaching conventional fully supervised competitors.