Pose-guided Visible Part Matching for Occluded Person ReID (2004.00230v1)

Abstract: Occluded person re-identification is a challenging task as the appearance varies substantially with various obstacles, especially in the crowd scenario. To address this issue, we propose a Pose-guided Visible Part Matching (PVPM) method that jointly learns the discriminative features with pose-guided attention and self-mines the part visibility in an end-to-end framework. Specifically, the proposed PVPM includes two key components: 1) pose-guided attention (PGA) method for part feature pooling that exploits more discriminative local features; 2) pose-guided visibility predictor (PVP) that estimates whether a part suffers the occlusion or not. As there are no ground truth training annotations for the occluded part, we turn to utilize the characteristic of part correspondence in positive pairs and self-mining the correspondence scores via graph matching. The generated correspondence scores are then utilized as pseudo-labels for visibility predictor (PVP). Experimental results on three reported occluded benchmarks show that the proposed method achieves competitive performance to state-of-the-art methods. The source codes are available at https://github.com/hh23333/PVPM

Pose-guided Visible Part Matching for Occluded Person ReID: An Overview

The paper “Pose-guided Visible Part Matching for Occluded Person ReID” presents a method specifically designed to tackle the occluded person re-identification (ReID) problem. This issue arises due to frequent occlusions in surveillance settings, which disrupt conventional ReID methods that assume the entire body of a person is visible. The authors address these challenges by proposing a Pose-guided Visible Part Matching (PVPM) method that leverages pose information to enhance feature discrimination and visibility estimation.

Key Components and Method

The PVPM method consists of several components critical to its operation:

1. Pose-guided Attention (PGA): This component is utilized for part feature pooling and aims to harness pose-guided attention to derive more discriminative local features. These attention masks are generated through a pose encoder that ingests pose-related information such as keypoint heatmaps and part affinity fields. The attention mechanism focuses on non-overlapping regions to extract complementary features from specific body parts.
2. Pose-guided Visibility Predictor (PVP): Part visibility prediction is crucial for matching only visible parts across images and thus mitigating occlusions. The PVP is trained in a self-supervised manner using pseudo-labels generated through graph matching mechanisms that establish part feature correspondences between positive pairs.
3. Graph Matching for Pseudo-label Generation: Visibility scores are generated as pseudo-labels by solving a feature correspondence problem using graph matching. This allows the network to self-mine part visibility rather than relying on biased external annotations.

Experimental Results

The evaluation of PVPM was conducted using several datasets including Occluded-REID, Partial-REID, and P-DukeMTMC-reID. The method demonstrated superior performance compared to existing holistic and occlusion-specific methods:

• Occluded-REID and Partial-REID: PVPM achieved competitive rank-1 accuracy, showcasing its ability to more effectively handle occluded samples by focusing on the visible parts guided by pose information.
• P-DukeMTMC-reID: Under both transfer learning and supervised settings, PVPM surpassed baseline methods, highlighting its robustness and adaptability to large-scale datasets.

Implications and Future Directions

The PVPM approach underscores the importance of leveraging structural information (such as pose) in tackling occluded ReID problems. The use of graph matching to derive visibility scores without external annotations is particularly indicative of a move towards more autonomous learning systems that could adapt to various ReID scenarios with minimal human intervention.

Looking forward, several speculative advancements could build upon this work:

• Integration with Real-time Systems: Implementing pose-guided visible part matching in real-time surveillance systems could be explored for automatic person tracking and identification in crowded, dynamic environments.
• Refinement of Graph Matching Techniques: Further development of graph matching algorithms could enhance the generation of pseudo-labels, especially in complex scenarios where occlusion patterns become substantially varied.
• Expansion to Multi-view Systems: Adapting PVPM to multi-camera systems might improve ReID performance by incorporating redundant visual information, potentially leading to more robust visibility prediction.

In conclusion, this paper contributes to the body of knowledge in occluded person ReID by providing a detailed methodology and demonstrating its effectiveness through empirical results, presenting PVPM as a promising approach for handling occlusions.