Part-Attention Based Model Make Occluded Person Re-Identification Stronger

TL;DR

This work tackles occluded person re-identification by introducing PAB-ReID, a part-attention based framework that leverages human parsing labels to generate precise body-part attention maps. It combines a Part Attention Block with a Global-local Learning Block, featuring a Pixel-level Attention Predictor, a Fine-Grained Feature Focuser, and a Part Triplet Loss to robustly align intra- and inter-class part features while suppressing background clutter. The model achieves state-of-the-art results on occluded datasets (Occluded-Duke, Occluded-reID, P-DukeMTMC) and strong performance on standard ReID benchmarks (Market-1501, DukeMTMC-ReID), significantly improving Rank-1 and mAP metrics. The proposed approach demonstrates the practical impact of integrating human semantic guidance, granular part features, and suffix losses to address occlusion and misalignment in ReID systems.

Abstract

The goal of occluded person re-identification (ReID) is to retrieve specific pedestrians in occluded situations. However, occluded person ReID still suffers from background clutter and low-quality local feature representations, which limits model performance. In our research, we introduce a new framework called PAB-ReID, which is a novel ReID model incorporating part-attention mechanisms to tackle the aforementioned issues effectively. Firstly, we introduce the human parsing label to guide the generation of more accurate human part attention maps. In addition, we propose a fine-grained feature focuser for generating fine-grained human local feature representations while suppressing background interference. Moreover, We also design a part triplet loss to supervise the learning of human local features, which optimizes intra/inter-class distance. We conducted extensive experiments on specialized occlusion and regular ReID datasets, showcasing that our approach outperforms the existing state-of-the-art methods.

Figures (6)

• Figure 1: Examples of Challenges in Occluded Person Re-identification. In Figure. 1(a), similar occlusions reduce the gap between different categories. In Figure. 1(b), different occlusions can lead to an increase in distances within the same category. Figure. 1(c) illustrates that the same body part may have a similar appearance across different individuals.
• Figure 2: The detailed structure of PAB-ReID is shown in the figure above. The PAB-ReID model comprises a part attention block responsible for generating part attention maps and a global-local learning block for extracting body part features. This paper generates six feature vectors representing the head, left hand, right hand, forehead, left leg, and right leg. In the above figure, the pedestrian's legs are occluded, and the visibility of the occluded leg is set to 0 when calculating the visibility score.
• Figure 3: The detail architecture of pixel-level attention predictor.
• Figure 4: The detail architecture of fine-grained feature focuser is shown in the figure above.
• Figure 5: Ablation experiments of loss function. The first column represents ID loss, the second column represents triplet loss, the third column represents the sum of ID and triplet loss, and the fourth column represents part triplet loss. From the above figure, it can be observed that part triplet loss can better guide the model learning.