A semi-automated system for person re-identification adaptation to cross-outfit and cross-posture scenarios

Person re-identification (ReID) algorithms are often trained on multi-camera snapshots of individuals taken on the same day, wearing the same outfits. Models trained with such protocols often fail in many long-term, indoor applications where person matching must be done across days, necessitating that algorithms be able to adapt to changing clothing and body postures. This study presents a simple, yet effective, system to overcome this challenge in realistic settings. We collected a new dataset capturing the natural variations of office worker appearances across days. To teach a ReID algorithm to adapt, we designed a semi-automated identity labeling system that requires only a small set of identification inputs from human labelers. The system utilized instance segmentation algorithms to detect people and one-shot video segmentation algorithms to track individuals across frames. Identified footages are then fed into the image repository to continually fine-tune the ReID network. These experiments demonstrate the applicability of our proposed method in helping the ReID algorithm overcome the challenges of varied clothing and postures. Our system improves the performance (measured by mAP) compared to pre-trained benchmark by 2.46% for the standard ReID condition, by 18.19% for cross-outfit re-identification, by 22.94% for cross-posture re-identification, and by 19.17% for the cross-posture and cross-outfit setting. As such, we anticipate this method may be beneficial towards the multitude of applications that utilize machine vision to automatically recognize human subjects.