Research map / Vision-language and multimodal / Human and embodied modeling
LLaVA-ReID: Selective Multi-image Questioner for Interactive 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.
Introduces interactive person re-identification (Inter-ReID), a dialogue-based retrieval task, along with the Interactive-PEDES dataset and LLaVA-ReID, a multi-image questioner with a learned candidate selector and looking-forward question supervision.
Problem
Text-based person re-identification assumes complete, single-shot descriptions, but real witness descriptions are often partial or vague; the paper targets iterative refinement of such descriptions through questioning.
Approach
An LMM questioner (LLaVA-OneVision) receives a hard-pass selected subset of representative gallery candidates plus dialogue history and is trained on questions chosen by a one-step looking-forward criterion that picks the question yielding the best target rank, with a Gumbel-top-k differentiable selector.
Key results
- On Interactive-PEDES after 5 rounds, LLaVA-ReID reaches 73.20 R@1 and BRI 0.719, versus 65.44 R@1 and BRI 0.849 for PlugIR.
- Integrating LLaVA-ReID into IRRA raises CUHK-PEDES R@1 from 73.38 to 78.51 and RSTPReid R@1 from 60.20 to 69.85.
- Integrating LLaVA-ReID into RDE raises ICFG-PEDES R@1 from 67.68 to 69.98 and RSTPReid R@1 from 65.35 to 72.00.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LLaVA-ReID instead of PlugIR for question generation | Interactive person ReID · Interactive-PEDES, Round 3 | R@1: 60.34 -> 63.96 | Helped |
| Use LLaVA-ReID instead of PlugIR for question generation | Interactive person ReID · Interactive-PEDES, Round 5 | R@1: 65.44 -> 73.20 | Helped |
| Use LLaVA-ReID rather than PlugIR | Interactive person ReID · Interactive-PEDES | BRI: 0.849 -> 0.719 | Helped |
| Add LLaVA-ReID interaction to IRRA | IRRA · CUHK-PEDES | R@1: 73.38 -> 78.51 | Helped |
| Add LLaVA-ReID interaction to RDE | RDE · CUHK-PEDES | R@1: 75.94 -> 79.39 | Helped |
| Add LLaVA-ReID interaction to IRRA | IRRA · ICFG-PEDES | R@1: 63.46 -> 67.44 | Helped |
| Add LLaVA-ReID interaction to RDE | RDE · ICFG-PEDES | R@1: 67.68 -> 69.98 | Helped |
| Add LLaVA-ReID interaction to IRRA | IRRA · RSTPReid | R@1: 60.20 -> 69.85 | Helped |
| Add LLaVA-ReID interaction to RDE | RDE · RSTPReid | R@1: 65.35 -> 72.00 | Helped |
| Use a learned selector instead of k-means candidate selection (ablation) | LLaVA-ReID · Interactive-PEDES, Round 5 | R@1: 72.51 -> 73.20 | Helped |
| Use a learned selector instead of uniform candidate sampling (ablation) | LLaVA-ReID · Interactive-PEDES, Round 5 | R@1: 72.70 -> 73.20 | Helped |
| Use a learned selector instead of top-k candidate selection (ablation) | LLaVA-ReID · Interactive-PEDES, Round 5 | R@1: 72.53 -> 73.20 | Helped |
Surprising findings
- Integrating the interactive module into T-ReID models yields gains even on ICFG-PEDES, which has human-annotated fine-grained descriptions, suggesting partial descriptions persist.
Limitations
- The retriever's limited ability to handle incremental, dialogue-based descriptions is acknowledged as a remaining challenge.
- Witness is simulated by an LLM answering from generated fine-grained descriptions rather than real humans.
Experimental setup
Datasets: CUHK-PEDES ICFG-PEDES Interactive-PEDES RSTPReid
Benchmarks: Interactive-PEDES CUHK-PEDES ICFG-PEDES RSTPReid
Metrics: Recall@1 Recall@5 Recall@10 mAP BRI
Baselines: SimIRV ChatIR PlugIR CFine RaSa APTM AUL IRRA RDE
Builds on: LLaVA IRRA PlugIR Gumbel-top-k
Closest papers by meaning
- Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification (CVPR 2025 · similarity 0.90)
- Multi-Prompts Learning with Cross-Modal Alignment for Attribute-Based Person Re-identification (AAAI 2024 · similarity 0.89)
- Instruct-ReID: A Multi-purpose Person Re-identification Task with Instructions (CVPR 2024 · similarity 0.88)
- ReID5o: Achieving Omni Multi-modal Person Re-identification in a Single Model (NeurIPS 2025 · similarity 0.87)
- Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment (CVPR 2025 · similarity 0.86)
- Modeling Thousands of Human Annotators for Generalizable Text-to-Image Person Re-identification (CVPR 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.