Hierarchical Multi-Graphs Learning for Robust Group Re-Identification

TL;DR

This work tackles the challenging problem of Group Re-identification (G-ReID) by introducing Hierarchical Multi-Graphs Learning (HMGL), which models a group as a set of multi-relational graphs capturing explicit cues (appearance, occlusion, foreground) and implicit dependencies. A dedicated Multi-Graphs Neural Network (MGNN) learns robust member representations by propagating information across these graphs, while Multi-Scale Matching (MSM) performs node-, subgraph-, and graph-level matching to mitigate hard-sample effects and intra-group ambiguity. Empirical results on RoadGroup and CUHK-SYSU Group establish state-of-the-art performance, with strong improvements in Rank-1 accuracy and mAP, and ablation studies confirm the value of each relational component and the reconstruction loss. The proposed framework enhances robustness to dynamic group structures and occlusions, offering a scalable approach for real-world multi-agent surveillance scenarios and potential extensions to broader group-reasoning tasks.

Abstract

Group Re-identification (G-ReID) faces greater complexity than individual Re-identification (ReID) due to challenges like mutual occlusion, dynamic member interactions, and evolving group structures. Prior graph-based approaches have aimed to capture these dynamics by modeling the group as a single topological structure. However, these methods struggle to generalize across diverse group compositions, as they fail to fully represent the multifaceted relationships within the group. In this study, we introduce a Hierarchical Multi-Graphs Learning (HMGL) framework to address these challenges. Our approach models the group as a collection of multi-relational graphs, leveraging both explicit features (such as occlusion, appearance, and foreground information) and implicit dependencies between members. This hierarchical representation, encoded via a Multi-Graphs Neural Network (MGNN), allows us to resolve ambiguities in member relationships, particularly in complex, densely populated scenes. To further enhance matching accuracy, we propose a Multi-Scale Matching (MSM) algorithm, which mitigates issues of member information ambiguity and sensitivity to hard samples, improving robustness in challenging scenarios. Our method achieves state-of-the-art performance on two standard benchmarks, CSG and RoadGroup, with Rank-1/mAP scores of 95.3%/94.4% and 93.9%/95.4%, respectively. These results mark notable improvements of 1.7% and 2.5% in Rank-1 accuracy over existing approaches.

Figures (10)

• Figure 1: Motivation illustration. Groups exhibit intricate relationships among members, including occlusion, foreground, appearance, and less obvious implicit connections. These densely intertwined relationships challenge the ability of a single graph to comprehensively represent the group structure, necessitating a more robust, multi-graphs approach.
• Figure 2: Hierarchical Multi-Graphs Learning Framework for Group Re-Identification. The framework models groups as multi-relational graphs, incorporating both explicit (e.g., foreground, occlusion, appearance) and implicit relationships to capture complex dependencies within group images. A reconstruction loss $\mathcal{L}^{re}$ is applied to enhance the representational capacity of these relationship-aware graphs. The Multi-Graphs Neural Network(MGNN) then explores the various relational dependencies within groups, yielding robust feature representations for each group member. Finally, group similarity is computed using the Multi-Scale Matching(MSM) approach, enabling effective comparison and re-identification across group instances.
• Figure 3: Comparative analysis of group re-identification methods based on Rank-1 to Rank-5 retrieval performance. Two query examples (a and b) are shown on the left for each method, with retrieved images ranked in descending order of similarity. Green-bordered images indicate correctly matched groups, while red-bordered images indicate incorrect matches.
• Figure 4: Visualization of various graph-based models for group re-identification, showcasing the structure and connectivity in each approach, alongside an analysis of feature map inconsistencies among group members. Each member is represented by an 8$\times$4 feature vector from the last convolutional layer, with a global average pooling vector representing individual identities. Ideally, each section of a member’s feature map should align uniquely with their own representation, but some regions display similarity to other members’ representations, indicating potential anomalies in the group re-identification process.
• Figure 5: Analysis of the impact of key parameters on the CMC at Rank-1, Rank-5, Rank-10, and Rank-20.