Mining Diverse Clues with Transformers for Person Re-identification

Abstract

Person Re-identification (ReID) is a challenging task due to the inherently large intra-class variation and subtle inter-class differences. Early works mainly tackle this problem by learning a discriminative pedestrian feature representation. Recently, vision transformer (ViT) has shown outstanding performance in many tasks, where the self-attention mechanism plays a key role that links every patch tokens with the class token. Intuitively, the class token could serve as a good global feature that captures local details naturally. Experiments demonstrate that the vanilla ViT can achieve impressive performance when directly applied to the ReID problem. Nevertheless, the class token may pay much attention to most salient patches while ignoring less salient but informative ones and missing some potential clues. To reduce this limitation, we propose a novel network MDCTNet to mine divers clues for person re-identification with transformers. Based on the cascaded architecture of transformer encoder, a Patch Suppression Module (PSM) is incorporated into the last several transformer layers, which aims to progressively discard some most salient patch tokens and make less salient ones passed to the next layer. Thus, the model is enforced to mine more potential useful clues in the remained patches and the resulting pedestrian representation can be more robust. Extensive experiments on three mainstream datasets including Market1501, DukeMTMC-ReID and MSMT17 validate the effectiveness of our method.