Improved Neighbor Feature Centralization for Person Re-Identification

Abstract

Person Re-Identification (ReID) aims to match pedestrian images across different camera views. Besides existing approaches that focus on designing effective descriptors for person representation, feature-level post-processing is considered as an efficient strategy for enhancing feature discriminability and improving ReID performance with low computational cost. Among these methods, Pose2ID is regarded as one of the most effective frameworks with the ability of learning discriminative cues from diverse-pose images for person representation. In this paper, we propose an improved Pose2ID framework with two main contributions. First, an adaptive weighting mechanism is adopted into the original framework to better reflect the role of each augmented image and mutual neighbor for person representation. Second, a cosine graph diffusion method is introduced to project extracted feature vectors into a manifold capturing the consistency among features belonging to the same identity. Experiments conducted on the Market-1501 and Market-1501 Partial datasets demonstrate the effectiveness of the proposed framework. The obtained results show that the proposed frame-work outperforms all compared methods in terms of mAP and Rank-1 ReID accuracy. Particularly, the proposed framework is highly effective on occluded datasets such as Market-1501 Partial, achieving improvements of 2.6% and 6.9% in mAP metric compared with the baseline Pose2ID framework with and without the IPG and NFC modules, respectively. The source code will be made publicly available soon.