PROJECT TITLE :

Pose-Guided Representation Learning for Person Re-Identification

ABSTRACT:

The high degree of difficulty in re-identifying a person is significantly increased by the large range of pose variations and misalignment errors that are displayed by person images (ReID). To alleviate those issues and improve the robustness of pedestrian representations, existing works frequently apply additional operations such as pose estimation and part segmentation, amongst other techniques. In spite of the fact that they improve ReID accuracy, the computational overheads introduced by those operations are significant, and the resulting deep models are difficult to fine-tune. In an effort to find a solution that is more effective, we have come up with the idea of a Part-Guided Representation, or PGR, which is made up of Pose Invariant Feature, or PIF, and Local Descriptive Feature, or LDF. Because local part cues are responsible for the training and supervision of PGR, we refer to it as "Part-Guided." In particular, the pose invariant function (PIF) is an approximation of a representation that is inferred by pose estimation and pose normalization. LDF places its primary emphasis on body parts that can be used for discrimination by approximating a representation obtained through body region segmentation. Extra pose extraction is only used during the training stage to supervise the learning of PGR; however, it is not required during the testing stage for feature extraction. This is because extra pose extraction is only used during the training stage. PGR's competitive accuracy and efficiency has been demonstrated by extensive comparisons with recent works on five widely used datasets.