Research map / Label-efficient and robust learning / Recognition and long-tail
High-Order Structure Based Middle-Feature Learning for Visible-Infrared Person Re-identification
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 HOS-Net, which combines CNN/Transformer short- and long-range features, a whitened hypergraph module for high-order structure, and graph-attention middle features trained with an MRIC contrastive loss for visible-infrared person re-identification.
Problem
Visible-infrared person re-identification suffers from a large modality discrepancy, and existing feature-level methods ignore high-order structure among local features and directly minimize VIS-IR distances, making a reasonable common feature space hard to learn.
Approach
Uses a whitened hypergraph network to model high-order relations among local features while avoiding hypergraph model collapse, and generates middle features by graph-attention alignment of VIS/IR and short/long-range features, supervised by a modality-range identity-center contrastive loss.
Key results
- On SYSU-MM01 all-search, HOS-Net reaches 75.6 Rank-1 / 74.2 mAP, versus 70.6 / 66.9 for MMN and 74.7 / 71.8 for DEEN.
- On RegDB VIS-to-IR, HOS-Net reaches 94.7 Rank-1 / 90.4 mAP, above the second-best MMN at 91.6 Rank-1.
- Ablation on SYSU-MM01 shows the AGW baseline rising from 69.9 / 66.9 to 75.6 / 74.2 with SLE, HSL, CFL and MRIC combined.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SLE to the AGW baseline (ablation) | AGW on SYSU-MM01 and RegDB · SYSU-MM01 and RegDB | R-1 / mAP: 69.9 / 66.9 -> 71.7 / 69.4; 85.0 / 79.1 -> 89.6 / 84.8 | Helped |
| Add HSL to the model with SLE (ablation) | HOS-Net on SYSU-MM01 and RegDB · SYSU-MM01 and RegDB | R-1 / mAP: 71.7 / 69.4 -> 73.3 / 72.4; 89.6 / 84.8 -> 92.0 / 87.1 | Helped |
| Add CFL to the model with SLE and HSL (ablation) | HOS-Net on RegDB · RegDB | R-1 / mAP: 92.0 / 87.1 -> 92.7 / 87.8 | Helped |
| Add MRIC loss (ablation) | HOS-Net on SYSU-MM01 and RegDB · SYSU-MM01 and RegDB | R-1 / mAP: 74.0 / 72.9 -> 75.6 / 74.2; 92.7 / 87.8 -> 94.7 / 90.4 | Helped |
| Add a hypergraph without whitening (ablation) | Baseline+SLE on SYSU-MM01 · SYSU-MM01 | R-1 / mAP: 71.7 / 69.4 -> 72.5 / 70.3 | Helped |
| Add whitening to the hypergraph (ablation) | Hypergraph model on SYSU-MM01 and RegDB · SYSU-MM01 and RegDB | R-1 / mAP: 72.5 / 70.3 -> 73.3 / 72.4; 91.1 / 86.3 -> 92.0 / 87.1 | Helped |
| Generate middle features with GAT instead of addition or concatenation (ablation) | CFL on SYSU-MM01 · SYSU-MM01 | mAP: 72.9 | Helped |
| Use GAT to align both modality and range features rather than only one type (ablation) | CFL on SYSU-MM01 · SYSU-MM01 | R-1 / mAP: 73.4 / 72.3; 73.6 / 72.4 -> 74.0 / 72.9 | Helped |
| Add the intra-range MRIC term (ablation) | HOS-Net on SYSU-MM01 and RegDB · SYSU-MM01 and RegDB | R-1 / mAP: 74.0 / 72.9 -> 74.3 / 73.3; 92.7 / 87.8 -> 93.4 / 88.6 | Helped |
| Add the middle-feature MRIC term (ablation) | HOS-Net on SYSU-MM01 and RegDB · SYSU-MM01 and RegDB | R-1 / mAP: 74.0 / 72.9 -> 74.4 / 73.2; 92.7 / 87.8 -> 93.3 / 88.3 | Helped |
| Add both intra-range and middle-feature MRIC terms (ablation) | HOS-Net on SYSU-MM01 and RegDB · SYSU-MM01 and RegDB | R-1 / mAP: 75.0 / 73.8; 93.8 / 89.2 | Helped |
| Add the inter-modality MRIC term to the other MRIC terms (ablation) | HOS-Net on SYSU-MM01 and RegDB · SYSU-MM01 and RegDB | R-1 / mAP: 75.0 / 73.8 -> 75.6 / 74.2; 93.8 / 89.2 -> 94.7 / 90.4 | Helped |
Surprising findings
- The CFL module uses ground-truth labels to build middle features during training, so it is not used at test time; inference cost is the same with or without it.
- The training complexity is high (training time roughly 5-6x the baseline on SYSU-MM01), which the authors list as a limitation.
Limitations
- Training complexity is high, and the authors plan to reduce it.
- The CFL module needs ground-truth identity labels of VIS and IR images, so it cannot be used at test time.
- Increased FLOPs and parameter count relative to the AGW baseline.
Experimental setup
Datasets: SYSU-MM01 RegDB LLCM
Benchmarks: SYSU-MM01 RegDB LLCM
Metrics: R-1 mAP CMC
Baselines: D2RL Hi-CMD JSIA-ReID X-Modality DDAG LbA G2DA TSME SPOT PMT CAJ MMN
Builds on: AGW HGNN Vision Transformer triplet loss
Closest papers by meaning
- LVLM-Driven Attribute-Aware Modeling for Visible-Infrared Person Re-Identification (NeurIPS 2025 · similarity 0.90)
- Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification (CVPR 2024 · similarity 0.90)
- Learning Commonality, Divergence and Variety for Unsupervised Visible-Infrared Person Re-identification (NeurIPS 2024 · similarity 0.89)
- Empowering Visible-Infrared Person Re-Identification with Large Foundation Models (NeurIPS 2024 · similarity 0.89)
- Implicit Discriminative Knowledge Learning for Visible-Infrared Person Re-Identification (CVPR 2024 · similarity 0.89)
- TokenMatcher: Diverse Tokens Matching for Unsupervised Visible-Infrared Person Re-Identification (AAAI 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.