Research map / Label-efficient and robust learning / Recognition and long-tail
Implicit Discriminative Knowledge 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 IDKL, a dual-branch VI-ReID network that purifies modality-specific features (IN-guided information purifier) and distills their implicit identity-aware knowledge into modality-shared features via triplet graph structure alignment and class semantic alignment.
Problem
Visible-infrared person re-identification suffers from intra-class variation and cross-modal discrepancy; prior methods only use modality-shared features and discard discriminative cues present in modality-specific features.
Approach
Extract modality-specific and modality-shared features with a dual-branch ResNet-50 (discriminator vs. adversarial confuser), purify the specific features with an IN-guided dual-mask module so they lose style discrepancy but keep identity cues, then distill them into the shared branch at feature level (TGSA graph alignment) and logit level (CSA KL), plus a modality discrepancy reduction loss.
Key results
- On SYSU-MM01 all-search single-shot, IDKL reaches 81.42% rank-1 and 79.85% mAP, above the listed SOTA baselines.
- On RegDB, IDKL achieves 94.72% rank-1 (visible-to-infrared) and 94.22% (infrared-to-visible), with mAP 90.19% and 90.43%.
- Ablation on SYSU-MM01 shows the full model reaching 81.42% rank-1 versus 67.60% for the baseline with only Lb.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use IDKL instead of the strongest listed baseline, SAAI | SYSU-MM01 all-search single-shot · SYSU-MM01 all-search single-shot | rank-1: 75.90 -> 81.42 | Helped |
| use IDKL instead of the strongest listed baseline, SAAI | SYSU-MM01 all-search single-shot · SYSU-MM01 all-search single-shot | mAP: 77.03 -> 79.85 | Helped |
| use IDKL instead of SAAI | SYSU-MM01 indoor-search single-shot · SYSU-MM01 indoor-search single-shot | rank-1: 83.20 -> 87.14 | Helped |
| use IDKL instead of SAAI | SYSU-MM01 indoor-search single-shot · SYSU-MM01 indoor-search single-shot | mAP: 88.01 -> 89.37 | Helped |
| use IDKL instead of MUN | RegDB visible-to-infrared · RegDB visible-to-infrared | rank-1: 95.19 -> 94.72 | Hurt |
| use IDKL instead of SAAI | RegDB visible-to-infrared · RegDB visible-to-infrared | mAP: 91.45 -> 90.19 | Hurt |
| use IDKL instead of MUN | RegDB infrared-to-visible · RegDB infrared-to-visible | rank-1: 91.86 -> 94.22 | Helped |
| use IDKL instead of SAAI | RegDB infrared-to-visible · RegDB infrared-to-visible | mAP: 92.01 -> 90.43 | Hurt |
| use IDKL instead of DEEN | LLCM visible-to-infrared · LLCM visible-to-infrared | rank-1: 62.5 -> 72.22 | Helped |
| use IDKL instead of DEEN | LLCM visible-to-infrared · LLCM visible-to-infrared | mAP: 65.8 -> 66.43 | Helped |
| use IDKL instead of DEEN | LLCM infrared-to-visible · LLCM infrared-to-visible | rank-1: 54.9 -> 70.72 | Helped |
| use IDKL instead of DEEN | LLCM infrared-to-visible · LLCM infrared-to-visible | mAP: 62.9 -> 65.19 | Helped |
Limitations
- Hyper-parameters (lambda1, lambda2, lambda3) are tuned by sweeping on the benchmark, with no separate validation protocol described.
- Gains on RegDB visible-to-infrared rank-1 are not uniformly best (CMT at 95.17% and MUN at 95.19% are higher), and SYSU-MM01 indoor/multi-shot results are mixed.
Experimental setup
Datasets: SYSU-MM01 LLCM RegDB
Benchmarks: SYSU-MM01 RegDB LLCM
Metrics: rank-1 rank-10 rank-20 mAP
Baselines: DDAG NFS CMT DEEN MPANet PIC MSCLNet PartMix SAAI MUN
Builds on: AGW-style ResNet-50 ReID DDAG gradient reversal layer deep mutual learning
Closest papers by meaning
- Discover Cross-Modality Nuances for Visible-Infrared Person Re-Identification (CVPR 2021 · similarity 0.93)
- Empowering Visible-Infrared Person Re-Identification with Large Foundation Models (NeurIPS 2024 · similarity 0.92)
- Learning Progressive Modality-Shared Transformers for Effective Visible-Infrared Person Re-identification (AAAI 2023 · similarity 0.92)
- Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification (CVPR 2024 · similarity 0.91)
- TokenMatcher: Diverse Tokens Matching for Unsupervised Visible-Infrared Person Re-Identification (AAAI 2025 · similarity 0.89)
- High-Order Structure Based Middle-Feature Learning for Visible-Infrared Person Re-identification (AAAI 2024 · similarity 0.89)
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.