Feature-Tuning Hierarchical Transformer via token communication and sample aggregation constraint for object re-identification.

Abstract

Recently, transformer-based methods have shown remarkable success in object re-identification. However, most works directly embed off-the-shelf transformer backbones for feature extraction. These methods treat all patch tokens equally, ignoring the difference of distinct patch tokens for feature representation. To solve this issue, this paper designs a feature-tuning mechanism for transformer backbones to emphasize important patches and attenuate unimportant patches. Specifically, a Feature-tuning Hierarchical Transformer (FHTrans) for object re-identification is proposed. First, we propose a plug-and-play Feature-tuning module via Token Communication (TCF) deployed within transformer encoder blocks. This module regards the class token as a pivot to achieve communication between patch tokens. Important patch tokens are emphasized, while unimportant patch tokens are attenuated, focusing more precisely on the discriminative features related to object distinction. Then, we construct a FHTrans based on the designed feature-tuning module. The encoder blocks are divided into three hierarchies considering the correlation between feature representativeness and transformer depth. As the hierarchy deepens, the communication between tokens becomes tighter. This enables the model to capture more crucial feature information. Finally, we propose a Sample Aggregation (SA) loss to impose more effective constraints on statistical characteristics among samples, thereby enhancing intra-class aggregation and guiding FHTrans to learn more discriminative features. Experiments on object re-identification benchmarks demonstrate that our method can achieve state-of-the-art performance.