Research map / Graph learning / Knowledge and temporal graphs

Motif Guided Graph Transformers with Combinatorial Skeleton Prototype Learning for Skeleton-Based Person Re-Identification

AAAI 2025 · Graph learning › Knowledge and temporal graphs · approach: Graph reasoning for visual recognition · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MoCos, a motif-guided graph transformer (hierarchical structural and gait collaborative motifs) with combinatorial sub-skeleton/sub-tracklet prototype contrastive learning for skeleton-based person re-ID.

Problem

Skeleton-based person re-identification suffers from assuming virtual relations among all joints, using average joint/sequence representations, and not exploiting gait-related structure or combinatorial spatial-temporal sub-patterns.

Approach

Endows joints with structural and gait-collaborative motif roles to guide graph transformer attention (MGT), and contrasts randomly masked sub-skeleton and sub-tracklet representations against identity prototypes (CSP).

Key results

• On BIWI-S, MoCos reaches 32.1 mAP and 72.0 Rank-1, versus 30.1 mAP and 68.7 Rank-1 for TranSG.
• On KGBD, MoCos reaches 26.1 mAP and 62.0 Rank-1, improving over TranSG (20.2 mAP, 59.0 Rank-1).
• Adding MoCos to SimMC raises BIWI-S mAP from 12.3 to 16.0 and Rank-1 from 41.7 to 55.4 in the unsupervised setting.

Measured effects reported in the paper

Surprising findings

• On CASIA-B C-N (clothes probe, normal gallery), MoCos has only slightly higher or lower Rank-1 than Hi-MPC (26.5 vs 27.2 Rank-1 in the table's bolded value), i.e., gains are not uniform across all settings.

Limitations

• Gains over TranSG are modest on some datasets (e.g., KS20 R10 equal at 90.2).
• Hyperparameters such as masking probabilities and fusion coefficient are tuned per dataset.
• Motif definitions are tailored to skeleton joint topology and limb grouping.

Experimental setup

Datasets: IAS-A IAS-B KS20 BIWI-S BIWI-W KGBD CASIA-B

Benchmarks: BIWI-S BIWI-W KS20 IAS-A IAS-B KGBD CASIA-B

Metrics: mAP Rank-1 Rank-5 Rank-10

Baselines: PoseGait AGE SGELA SimMC Hi-MPC MG-SCR SM-SGE SPC-MGR ST-GCN TranSG D_PG D_13

Builds on: graph transformer SimMC TranSG PoseGait

Closest papers by meaning

• Revealing Key Details to See Differences: A Novel Prototypical Perspective for Skeleton-based Action Recognition (CVPR 2025 · similarity 0.87)
• BlockGCN: Redefine Topology Awareness for Skeleton-Based Action Recognition (CVPR 2024 · similarity 0.86)
• SkeletonGait: Gait Recognition Using Skeleton Maps (AAAI 2024 · similarity 0.86)
• Heterogeneous Skeleton-Based Action Representation Learning (CVPR 2025 · similarity 0.86)
• SAMoR: Motion Modelling for Articulated Objects of Any Skeleton and Topology (NeurIPS 2026 · similarity 0.85)
• Skeleton-based Action Recognition with Non-linear Dependency Modeling and Hilbert-Schmidt Independence Criterion (AAAI 2025 · similarity 0.85)