Inter-block ladder-style transformer model with multi-subspace feature adjustment for object re-identification

Abstract

Object re-identification aims to retrieve specific objects across multiple cameras and has garnered significant attention. Currently, transformer-based methods have taken a dominant position. However, most approaches embed inherent transformer encoders for feature extraction directly. These methods handle all patch tokens uniformly, failing to distinguish salient and non-salient patch tokens for discriminative feature expression. To this end, this work proposes a novel inter-block ladder-style transformer (IBLSFormer) for object re-identification. Firstly, a multi-subspace feature adjustment (MSFA) module is designed to adjust the patch features via class-patch interaction in multiple subspaces including Euclidean distance subspace, cosine distance subspace, and KL divergence subspace. The MSFA module can enhance the salient patch tokens and weaken the non-salient patch tokens simultaneously to focus on discriminative patches. Afterwards, an IBLSFormer is designed by inserting MSFA modules with distinct configurations into the vision transformer. The narrow-to-wide ladder-style constraints are embedded in MSFA modules based on embedding depth to highlight the feature differences across different levels and ameliorate the feature learning. Our method achieves mAP/Rank-1 of 88.7%/95.3%, 81.4%/90.4%, 80.0%/97.1%, and 89.4%/83.6% on four object re-identification datasets. Extensive experiments show IBLSFormer is superior to other methods in learning discriminative and robust representations for object re-identification.