Intra-Camera Supervised Person Re-ID by Tracklet Level Classifier

摘要

In this work, we propose a novel method to perform intra-camera supervised person re-ID by Tracklet Level Classifier (TLC). The key idea of our method is to train classifiers for every intra-camera ID, which is tracklet level, compared with camera level of previous works. By training tracklet level classifiers, we make the backbone learned to extract intra-camera invariant representations. With the fine-trained classifiers, we mine and exploit latent inter-camera ID matching pairs easily. Previous works needs two stages and relies on complicated rules to match inter-camera pairs while we simplify the training strategy to only one stage and do not need a complex design to match tracklet over cameras. Extensive experiments and ablation studies on three large re-ID datasets show that our simple and effective TLC method achieve state-of-the-art among all the intra-camera supervised person re-ID methods.