DLCR: A Generative Data Expansion Framework via Diffusion for Clothes-Changing Person Re-ID

With the recent exhibited strength of generative diffusion models, an openresearch question is if images generated by these models can be used to learnbetter visual representations. While this generative data expansion may sufficefor easier visual tasks, we explore its efficacy on a more difficultdiscriminative task: clothes-changing person re-identification (CC-ReID).CC-ReID aims to match people appearing in non-overlapping cameras, even whenthey change their clothes across cameras. Not only are current CC-ReID modelsconstrained by the limited diversity of clothing in current CC-ReID datasets,but generating additional data that retains important personal features foraccurate identification is a current challenge. To address this issue wepropose DLCR, a novel data expansion framework that leverages pre-traineddiffusion and large language models (LLMs) to accurately generate diverseimages of individuals in varied attire. We generate additional data for fivebenchmark CC-ReID datasets (PRCC, CCVID, LaST, VC-Clothes, and LTCC) andincrease their clothing diversity by 10X, totaling over 2.1M images generated.DLCR employs diffusion-based text-guided inpainting, conditioned on clothingprompts constructed using LLMs, to generate synthetic data that only modifies asubject's clothes while preserving their personally identifiable features. Withthis massive increase in data, we introduce two novel strategies - progressivelearning and test-time prediction refinement - that respectively reducetraining time and further boosts CC-ReID performance. On the PRCC dataset, weobtain a large top-1 accuracy improvement of 11.3% by training CAL, a previousstate of the art (SOTA) method, with DLCR-generated data. We publicly releaseour code and generated data for each dataset here:https://github.com/CroitoruAlin/dlcr.