| Summary: | Person Re-Identification, which has been extensively researched for its wide applicability, concentrates solely on the entity of “person” within image retrieval cases. This presents a significant challenge. The person re-identification task, which follows the detector algorithm step, is limited to information within the detected box. This limitation leads to what we refer to as “in-box noise,”, a type of noise that is detrimental to the predefined notion of the box-identity pair, caused by any object, other identity, or environment that obscures or interferes with the recognition of the target identity within the box. To address this in-box noise, we propose a methodology that involves training with a self-supervised teacher model. This approach exploits the relevant identity information within the box-identity pair, enabling parallel learning with the main re-identification task and interpreting the critical identity areas in the image as guided by the teacher. This methodology demonstrates impressive performance on benchmark datasets, achieving a mean average precision (mAP) of 73.4 and a rank-1 score of 88.8 on MSMT17, and an mAP of 89.7 and a rank-1 score of 95.2 on Market1501.
|
|---|
