AATransID : enhancing cross-domain person re-identification with low-rank adaptation and sequence sampling

Abstract

Person re-identification (ReID) is a pivotal task in computer vision, aiming to match individuals across non-overlapping camera views. Despite significant advancements, current ReID methodologies often suffer from limited generalization capabilities, especially in cross-domain scenarios where training and testing datasets differ in environments, camera settings, and appearance variations. To address these challenges, we propose AATransID, a novel person ReID model based on a pure Vision Transformer architecture. We introduce a sequence sampling methodology that acts as an effective regularization technique, exposing the model to a wide range of intra-class variations and temporal contexts, thereby enhancing its robustness. Additionally, we integrate ArcFace loss with triplet loss to refine the feature space. With low-rank adaptation, AATransID achieves substantial improvements in mean average precision (mAP) on multiple benchmark datasets, surpassing existing state-of-the-art models by 4.8% on Market-1501, 1.6% on DukeMTMC-reID, 7.7% on MSMT17, and 2.2% on Occluded-Duke. Source code is available at https://github.com/serdaryildiz/AATransID.