Unsupervised Graph Association for Person Re-Identification

In this paper, we propose a novel unsupervised graph association (UGA) to learn the underlying view-invariant representations from the video pedestrian tracklets. The core points of it are mining the cross-view relationships and reducing the damage of noisy associations. To this end, UGA adopts a two-stage training strategy: (1) intra-camera learning stage and (2) inter-camera learning stage. The former is to learn representations of a person with regards to camera information, which helps to reduce false crossview associations in the second stage. Compared with existing tracklet-based methods, ours can build more accurate cross-view associations and require lower GPU memory. Extensive experiments and ablation studies on seven RE-ID datasets demonstrate the superiority of the proposed UGA over most state-of-the-art unsupervised and domain adaptation RE-ID methods. Code is available at github 1 .