Research map / Vision-language and multimodal / Human and embodied modeling
Multi-Prompts Learning with Cross-Modal Alignment for Attribute-Based 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 MP-ReID, which combines explicit prompts (ChatGPT-generated and VQA-generated attribute sentences) with learnable implicit prompts on a CLIP backbone, aligned via cross-attention and a multi-modal Transformer, to improve attribute-based person re-identification on Market1501 and DukeMTMC-reID.
Problem
Person re-identification currently under-uses fine-grained attribute descriptions, relying on coarse attribute words; there is also a cross-modal gap between attribute text prompts and person images, and no prompt-related ReID dataset exists at scale.
Approach
Fuse multiple prompt types (explicit ChatGPT and VQA attribute sentences plus CoOp-style learnable implicit prompts) with cross-attention and a multi-modal Transformer, trained with classification, contrastive, and triplet/ID losses to align prompts with images.
Key results
- MP-ReID reaches 95.50 mAP and 97.70 R@1 on Market1501, versus 89.60 mAP and 95.50 R@1 for CLIP-ReID.
- On DukeMTMC-reID, MP-ReID reaches 88.90 mAP and 95.70 R@1, versus 82.50 mAP and 90.00 R@1 for CLIP-ReID.
- Ablation shows the combined ChatGPT+VQA explicit prompts with learnable prompts give the best results; removing VQA prompts drops Market1501 mAP by 7.9 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MP-ReID with multiple fine-grained prompts instead of CLIP-ReID | Person ReID · Market1501 | mAP: 89.60 -> 95.50 | Helped |
| use MP-ReID with multiple fine-grained prompts instead of CLIP-ReID | Person ReID · DukeMTMC-reID | mAP: 82.50 -> 88.90 | Helped |
| add coarse and separate attribute words (ablation) | CLIP-ReID learnable-prompt baseline · Market1501 | mAP: 89.60 -> 89.40 | Hurt |
| add coarse and separate attribute words (ablation) | CLIP-ReID learnable-prompt baseline · DukeMTMC-reID | mAP: 82.50 -> 83.90 | Helped |
| add generation caption prompts (ablation) | CLIP-ReID learnable-prompt baseline · Market1501 | mAP: 89.60 -> 86.30 | Hurt |
| add generation caption prompts (ablation) | CLIP-ReID learnable-prompt baseline · DukeMTMC-reID | mAP: 82.50 -> 78.10 | Hurt |
| add VQA prompts (ablation) | CLIP-ReID learnable-prompt baseline · Market1501 | mAP: 89.60 -> 87.60 | Hurt |
| add VQA prompts (ablation) | CLIP-ReID learnable-prompt baseline · DukeMTMC-reID | mAP: 82.50 -> 78.20 | Hurt |
| add ChatGPT prompts (ablation) | CLIP-ReID learnable-prompt baseline · Market1501 | mAP: 89.60 -> 90.20 | Helped |
| add ChatGPT prompts (ablation) | CLIP-ReID learnable-prompt baseline · DukeMTMC-reID | mAP: 82.50 -> 87.20 | Helped |
| add both ChatGPT and VQA prompts (ablation) | CLIP-ReID learnable-prompt baseline · Market1501 | mAP: 89.60 -> 95.50 | Helped |
| add both ChatGPT and VQA prompts (ablation) | CLIP-ReID learnable-prompt baseline · DukeMTMC-reID | mAP: 82.50 -> 88.90 | Helped |
Surprising findings
- Coarse attribute words alone with learnable prompts slightly lowered Market1501 mAP versus baseline in some configurations, and generation captions performed worse than ChatGPT prompts.
Limitations
- Only three prompt types are explored; the integration strategy is relatively simple.
- Multi-modal large models were not extensively evaluated due to data and resource constraints.
- Evaluated only on two datasets with manually annotated attributes.
Experimental setup
Datasets: Market1501 DukeMTMC-reID
Benchmarks: Market1501 DukeMTMC-reID
Metrics: mAP R@1 R@5 R@10
Baselines: SAN PAT TransReID MSDPA DCAL AANet AMD UCAD UPAR CLIP-ReID
Builds on: CLIP-ReID CoOp CLIP
Closest papers by meaning
- ReID5o: Achieving Omni Multi-modal Person Re-identification in a Single Model (NeurIPS 2025 · similarity 0.89)
- LLaVA-ReID: Selective Multi-image Questioner for Interactive Person Re-Identification (ICML 2025 · similarity 0.89)
- Modeling Thousands of Human Annotators for Generalizable Text-to-Image Person Re-identification (CVPR 2025 · similarity 0.89)
- Instruct-ReID: A Multi-purpose Person Re-identification Task with Instructions (CVPR 2024 · similarity 0.88)
- Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification (CVPR 2025 · similarity 0.87)
- Unifying Multi-Modal Uncertainty Modeling and Semantic Alignment for Text-to-Image Person Re-identification (AAAI 2024 · similarity 0.87)
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.