Boosting occluded person re-identification by leveraging occlusion attributes

Abstract

Occluded person re-identification (ReID) aims to address the potential occlusion problem when matching occluded or holistic pedestrians from different camera views. Currently, occlusion augmentation-based methods have not fully exploited the occlusion attributes, resulting in suboptimal results. We introduce a novel ReID framework, dubbed Occlusion Attributes boosted Occluded Person Re-Identification (OA-ReID), aimed at leveraging the occlusion attributes for pedestrian-focused feature learning. Firstly, we propose an occlusion emulator (OE) that generates artificially occluded images towards emulating the occlusion scenarios. Both the original image and the corresponding artificially occluded image are jointly used for model training. Secondly, we present two crucial components, namely the inductive hard (IH) sample mining and the Occlusion-Informed Part Transformer (OIPT). The IH sample mining leverages the obstacle category to construct inductive triplets, which induces the model to extract identity-relevant features. The OIPT integrates the obstacle position information into our ReID framework to rectify the erroneous attention on occlusions, promoting reliable target pedestrian localization. Through extensive experiments, we show OA-ReID achieves state-of-the-art performance on both occluded and holistic person ReID benchmarks.