Attribute‐guided transformer for robust person re‐identification

Recent studies reveal the crucial role of local features in learning robust and discriminative representations for person re-identification (Re-ID). Existing approaches typically rely on external tasks, example, semantic segmentation, or pose estimation, to locate identifiable parts given images. However, they heuristically utilise predictions from off-the-shelf models, which may be sub-optimal terms both partition computational efficiency. They also ignore mutual information with other inputs, weakens representation capabilities features. In this study, authors put forward a novel Attribute-guided Transformer (AiT), explicitly exploits pedestrian attributes as priors learning. Specifically, first introduce an attribute process, generates set attention maps highlighting informative Then, design Feature Diffusion Module (FDM) iteratively inject into global feature maps, aiming at suppressing unnecessary noise inferring attribute-aware representations. Last, propose Aggregation (FAM) exploit aggregating characteristics different images, enhancing embedding. Extensive experiments demonstrate superiority our AiT As result, achieve competitive performance state-of-the-art methods several challenging benchmarks without any bells whistles.