Transformer Meets Part Model: Adaptive Part Division for Person Re-Identification

Shenqi Lai, Zhenhua Chai, Xiaolin Wei; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021, pp. 4150-4157

Abstract


Part model is one of the key factors to high performance person re-identification (ReID) task. In recent studies, there are mainly two streams for part model. The first one is to divide a person image into several fixed parts to obtain their local information, but it may cause performance degradation in case of misalignment. The other one is to explore external resources like pose estimation or human parsing to locate local parts, but it costs extra storage and computation. Inspired by recent successful transformers on spatial similarity modeling, we propose a novel Adaptive Part Division (APD) model to better extract local features. More specifically, APD mainly consists of two crucial modules: a Transformer-based Part Merge (TPM) module and a Part Mask Generation (PMG) module. In particular, TPM first adaptively assigns the patch tokens of the same semantic object to the identical part. Then, PMG takes these identical parts together and generates several non-overlapping masks for robust part division. We have conducted extensive evaluations on four popular benchmarks, i.e. Market-1501, CUHK03, DukeMTMC-ReID and MSMT17, and the experimental results show that our proposed method achieves the state-of-the-art performance.

Related Material


[pdf]
[bibtex]
@InProceedings{Lai_2021_ICCV, author = {Lai, Shenqi and Chai, Zhenhua and Wei, Xiaolin}, title = {Transformer Meets Part Model: Adaptive Part Division for Person Re-Identification}, booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops}, month = {October}, year = {2021}, pages = {4150-4157} }