Research map / Vision-language and multimodal / Cross-modal learning
UGG-ReID: Uncertainty-Guided Graph Model for Multi-Modal Object 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 UGG-ReID, a CLIP-based multi-modal object re-identification method that models local patch and global aleatoric uncertainty with a Gaussian patch-graph and routes samples through an uncertainty-guided mixture of experts.
Problem
Multi-modal object ReID (RGB, NIR, TIR) is hurt by fine-grained local noise such as occlusion and frame loss, and by inconsistent modality quality, which makes fusion unstable.
Approach
Represents patch and global tokens as Gaussian distributions connected in a graph (GPGR) and uses sample-level uncertainty to gate a mixture of experts (UGMoE) with uncertainty-guided routing for cross-modal interaction.
Key results
- On RGBNT201, UGG-ReID reaches 81.2 mAP and 86.8 R-1, versus 79.7 mAP / 81.8 R-1 for DeMo.
- On WMVEID863, UGG-ReID reaches 72.6 mAP and 80.8 R-1, outperforming the second best by 2.8 mAP and 3.6 R-1.
- On Market1501-MM, UGG-ReID reports 85.4 mAP and 94.3 R-1; it runs at 371.4 FPS with 103.2M params on RGBNT201.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add UGG-ReID instead of IDEA | CLIP-based ReID · RGBNT201 | mAP: 80.2 -> 81.2 | Helped |
| add UGG-ReID instead of MambaPro | CLIP-based ReID · Market1501-MM | mAP: 84.1 -> 85.4 | Helped |
| add UGG-ReID instead of ICPL-ReID | CLIP-based ReID · MSVR310 | mAP: 56.9 -> 60.1 | Helped |
| add UGG-ReID instead of FACENeT | multi-modal vehicle ReID · WMVEID863 | mAP: 69.8 -> 72.6 | Helped |
| add traditional MoE to the baseline (ablation) | CLIP baseline · RGBNT201 | mAP: 72.2 -> 75.0 | Helped |
| add uncertainty modeling to MoE (ablation) | UGMoE + baseline · RGBNT201 and WMVEID863 | mAP: 75.0 -> 76.0; 68.7 -> 69.7 | Helped |
| add standard PGR without uncertainty (ablation) | UGMoE with uncertainty · RGBNT201 and WMVEID863 | mAP: 76.0 -> 77.3; 69.7 -> 70.2 | Helped |
| add uncertainty to PGR (GPGR) (ablation) | UGMoE + PGR · RGBNT201 and WMVEID863 | mAP: 77.3 -> 81.2; 70.2 -> 72.6 | Helped |
| add the global-token constraint L_c,s (ablation) | UGG-ReID · RGBNT201 and WMVEID863 | mAP: 78.8 -> 81.2; 69.6 -> 72.6 | Helped |
| add the expert uncertainty loss L_r (ablation) | UGG-ReID · RGBNT201 and WMVEID863 | mAP: 79.0 -> 81.2; 70.6 -> 72.6 | Helped |
| add the expert-balancing loss L_e (ablation) | UGG-ReID · RGBNT201 and WMVEID863 | mAP: 81.0 -> 81.2; 71.2 -> 72.6 | Helped |
| use 4 experts rather than 2, 3, 5, or 6 (ablation) | UGG-ReID · RGBNT201 and WMVEID863 | mAP: 80.3; 80.8; 81.2; 80.0; 80.1; 71.4; 72.0; 72.6; 72.0; 71.1 | Mixed |
Surprising findings
- The paper's appendix describes the uncertainty as epistemic in its discussion, while the main text describes it as aleatoric.
Limitations
- May still struggle when local cues are heavily corrupted or missing, as the authors note.
- Gains on some datasets are modest, e.g. the Market1501-MM mAP is below MambaPro-level changes in some metrics and ablations rely on single runs.
- Hyperparameters such as the number of experts and nodes were tuned on the test-reported benchmarks.
Experimental setup
Datasets: RGBNT201 Market1501-MM MSVR310 RGBNT100 WMVEID863
Benchmarks: RGBNT201 Market1501-MM MSVR310 RGBNT100 WMVEID863
Metrics: mAP R-1 R-5 R-10 params FLOPs FPS
Baselines: HAMNet PFNet IEEE TIENet UniCat EDITOR RSCNet HTT TOP-ReID ICPL-ReID PromptMA MambaPro
Builds on: CLIP DeMo mixture of experts graph convolutional network
Closest papers by meaning
- Magic Tokens: Select Diverse Tokens for Multi-modal Object Re-Identification (CVPR 2024 · similarity 0.88)
- DeMo: Decoupled Feature-Based Mixture of Experts for Multi-Modal Object Re-Identification (AAAI 2025 · similarity 0.88)
- TOP-ReID: Multi-Spectral Object Re-identification with Token Permutation (AAAI 2024 · similarity 0.87)
- Multi-Modal Object Re-identification via Sparse Mixture-of-Experts (ICML 2025 · similarity 0.86)
- MambaPro: Multi-Modal Object Re-identification with Mamba Aggregation and Synergistic Prompt (AAAI 2025 · similarity 0.86)
- Cocoon: Robust Multi-Modal Perception with Uncertainty-Aware Sensor Fusion (ICLR 2025 · similarity 0.85)
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.