Unsupervised Domain Adaptation in Person re-ID via k-Reciprocal Clustering and Large-Scale Heterogeneous Environment Synthesis

An ongoing major challenge in computer vision is the task of person\nre-identification, where the goal is to match individuals across different,\nnon-overlapping camera views. While recent success has been achieved via\nsupervised learning using deep neural networks, such methods have limited\nwidespread adoption due to the need for large-scale, customized data\nannotation. As such, there has been a recent focus on unsupervised learning\napproaches to mitigate the data annotation issue; however, current approaches\nin literature have limited performance compared to supervised learning\napproaches as well as limited applicability for adoption in new environments.\nIn this paper, we address the aforementioned challenges faced in person\nre-identification for real-world, practical scenarios by introducing a novel,\nunsupervised domain adaptation approach for person re-identification. This is\naccomplished through the introduction of: i) k-reciprocal tracklet Clustering\nfor Unsupervised Domain Adaptation (ktCUDA) (for pseudo-label generation on\ntarget domain), and ii) Synthesized Heterogeneous RE-id Domain (SHRED) composed\nof large-scale heterogeneous independent source environments (for improving\nrobustness and adaptability to a wide diversity of target environments).\nExperimental results across four different image and video benchmark datasets\nshow that the proposed ktCUDA and SHRED approach achieves an average\nimprovement of +5.7 mAP in re-identification performance when compared to\nexisting state-of-the-art methods, as well as demonstrate better adaptability\nto different types of environments.\n