Research map / Vision-language and multimodal / Cross-modal learning
Adaptive Uncertainty-Based Learning for Text-Based Person Retrieval
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 Adaptive Uncertainty-based Learning (AUL), using Subjective Logic-based matching uncertainty to filter unreliable pairs, Gaussian-noise uncertainty representations for progressive coarse-to-fine alignment, and cross-modal masked modeling for text-based person retrieval.
Problem
Text-based person retrieval suffers from unreliable matching pairs (matching ambiguity) under large intra-class and small inter-class variation, and one-sided one-to-one cross-modal alignment that ignores one-to-many correspondence.
Approach
Model matching uncertainty with Subjective Logic to adaptively weight training pairs, build Gaussian-noise uncertainty representations to simulate coarse-grained one-to-many alignment, progressively blend coarse and fine alignment losses, and add cross-modal masked image/language modeling.
Key results
- On RSTPReid, AUL reaches 71.65% R@1 in the supervised setting, beating APTM (66.45%) and RaSa (66.90%).
- On CUHK-PEDES, AUL achieves 77.23% R@1, versus 76.51% for RaSa (the prior best in Table 1).
- In the domain generalization setting (ICFG-PEDES to CUHK-PEDES), AUL reaches 56.79% R@1 vs 48.67% for APTM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add UMF (ablation) | AUL · RSTPReid | R@1: 68.15 -> 69.55 | Helped |
| add UAR with both Lfa and Lca (ablation) | AUL · RSTPReid | R@1: 68.15 -> 69.15 | Helped |
| add CMM (ablation) | AUL · RSTPReid | R@1: 68.15 -> 69.25 | Helped |
| add CMM to UAR using both Lfa and Lca (ablation) | AUL · RSTPReid | R@1: 69.15 -> 70.95 | Helped |
| use both Lfa and Lca in UAR instead of Lfa only (ablation) | AUL · RSTPReid | R@1: 70.85 -> 70.95 | Helped |
| add UMF to UAR and CMM (ablation) | AUL · RSTPReid | R@1: 70.95 -> 71.65 | Helped |
| add CMM to UMF (ablation) | AUL · RSTPReid | R@1: 69.55 -> 71.35 | Helped |
| use progressive alignment weights γ rather than average weights (ablation) | AUL UAR · RSTPReid | retrieval accuracy | Helped |
| use easy-to-hard γ rather than reverse-weight ˆγ (ablation) | AUL UAR · RSTPReid | retrieval accuracy | Helped |
| use both CMLM and CMIM instead of CMLM alone (ablation) | AUL CMM · ICFG-PEDES and RSTPReid | retrieval performance | Helped |
| use AUL instead of APTM | text-based person retrieval · RSTPReid supervised | R@1: 66.90 -> 71.65 | Helped |
| use AUL instead of APTM | text-based person retrieval · ICFG-PEDES supervised | R@1: 68.22 -> 69.16 | Helped |
Surprising findings
- Applying the CMLM loss alone is less effective than the combined CMIM and CMLM (CMM) objective, suggesting both masked token types are needed.
Limitations
- Relies on a two-stage pretraining on the APTM dataset, so gains are coupled to that external data.
- Uncertainty is derived from similarity-based evidence, which the paper does not validate beyond ablation/distribution analysis.
Experimental setup
Datasets: CUHK-PEDES ICFG-PEDES RSTPReid APTM pretraining dataset
Benchmarks: CUHK-PEDES ICFG-PEDES RSTPReid
Metrics: Rank-1 Rank-5 Rank-10
Baselines: APTM RaSa IRRA CFine TP-TPS ASAMN C2A2 LGUR SRCF ISANet
Builds on: APTM MAE Subjective Logic InfoNCE
Closest papers by meaning
- Exploring the Potential of Large Vision-Language Models for Unsupervised Text-Based Person Retrieval (AAAI 2025 · similarity 0.91)
- Unifying Multi-Modal Uncertainty Modeling and Semantic Alignment for Text-to-Image Person Re-identification (AAAI 2024 · similarity 0.89)
- Automatic Synthetic Data and Fine-grained Adaptive Feature Alignment for Composed Person Retrieval (NeurIPS 2025 · similarity 0.88)
- Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment (CVPR 2025 · similarity 0.87)
- Causality-Inspired Invariant Representation Learning for Text-Based Person Retrieval (AAAI 2024 · similarity 0.87)
- DM-Adapter: Domain-Aware Mixture-of-Adapters for Text-Based Person Retrieval (AAAI 2025 · similarity 0.86)
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.