Research map / Vision-language and multimodal / CLIP and zero-shot
CLIP-driven View-aware Prompt Learning for Unsupervised Vehicle 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 ViewCoOp, a view-aware learnable textual prompt with dynamic view-ratio embeddings, and a cross-modal mutual graph co-matching loss to fine-tune CLIP for unsupervised vehicle Re-ID.
Problem
Unsupervised vehicle re-identification lacks textual descriptions that capture view variations; ID-only CLIP prompts (CoOp) cannot describe vehicle views, and instance-level image-text alignment ignores relations between samples.
Approach
A learnable prompt whose tokens embed per-view patch-level proportions (from a view-parsing model) plus dynamic learnable view biases, combined with a graph over samples whose edges come from bimodal (visual and textual) DBSCAN cluster agreement, used to define positive/negative pairs for a graph-aware loss while fine-tuning both CLIP encoders.
Key results
- On VeRi-776 the full method reaches 88.3% Rank-1 and 42.4% mAP, versus 80.3% Rank-1 and 36.8% mAP for the baseline (Table 1).
- On VehicleID Test2400 it reaches 61.3% Rank-1 and 47.4% mAP, versus 56.0% Rank-1 and 41.5% mAP for the baseline.
- Dynamic ViewCoOp improves over CoCoOp by 2.7 Rank-1 and 2.1 mAP points on VeRi-776 (Table 2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ViewCoOp (ablation) | CLIP-driven vehicle Re-ID baseline · VeRi-776 | Rank1: 80.3 -> 83.6 | Helped |
| Add cross-modal mutual graph co-matching (ablation) | CLIP-driven vehicle Re-ID baseline · VeRi-776 | Rank1: 80.3 -> 84.3 | Helped |
| Add both ViewCoOp and cross-modal mutual graph co-matching (ablation) | CLIP-driven vehicle Re-ID baseline · VeRi-776 | Rank1: 80.3 -> 88.3 | Helped |
| Add ViewCoOp (ablation) | CLIP-driven vehicle Re-ID baseline · VehicleID Test2400 | Rank1: 56.0 -> 58.6 | Helped |
| Add cross-modal mutual graph co-matching (ablation) | CLIP-driven vehicle Re-ID baseline · VehicleID Test2400 | Rank1: 56.0 -> 58.8 | Helped |
| Use dynamic view bias rather than static view proportions (ablation) | ViewCoOp · VeRi-776 | Rank1: 86.6 -> 88.3 | Helped |
| Use dynamic view bias rather than static view proportions (ablation) | ViewCoOp · VeRi-776 | mAP: 41.0 -> 42.4 | Helped |
| Set graph edge threshold τ to 1 or 2 instead of τ = 0 (ablation) | Cross-modal mutual graph co-matching · VeRi-776 | Rank1: 83.2 -> 88.3 | Helped |
| Set graph edge threshold τ to 1 or 2 instead of τ = 2 (ablation) | Cross-modal mutual graph co-matching · VeRi-776 | Rank1: 86.2 -> 88.3 | Helped |
| Calculate view ratios using 16 × 16 patches rather than pixel-level masks (ablation) | ViewCoOp · VeRi-776 | Rank1: 86.2 -> 88.3 | Helped |
| Calculate view ratios using 16 × 16 patches rather than 8 × 8 patches (ablation) | ViewCoOp · VeRi-776 | Rank1: 86.0 -> 88.3 | Helped |
| Calculate view ratios using 16 × 16 patches rather than 32 × 32 patches (ablation) | ViewCoOp · VeRi-776 | Rank1: 87.4 -> 88.3 | Helped |
Surprising findings
- Static view-proportion prompts (ViewCoOp Static) underperform on VehicleID Test2400 relative to UnCoOp (59.7 vs 60.6 Rank-1).
Limitations
- Pixel-level view ratios depend on segmentation model accuracy and degrade performance.
- Graph edge threshold is sensitive; using only single-modal or only bimodal constraints performs worse.
- Evaluated only on two vehicle Re-ID datasets, with no results reported for other multi-view cross-modal tasks.
Experimental setup
Datasets: VeRi-776 VehicleID
Benchmarks: VeRi-776 VehicleID
Metrics: mAP Rank-1 Rank-5
Baselines: MMT SPCL ICE GCMT ISE CTFRN PPLR Cluster Contrast AdaMG UCF CoOp CoCoOp
Builds on: CoOp CLIP-ReID CoCoOp PVEN
Closest papers by meaning
- Infer the Whole from a Glimpse of a Part: Keypoint-Based Knowledge Graph for Vehicle Re-Identification (AAAI 2025 · similarity 0.87)
- Structural Information Guided Multimodal Pre-training for Vehicle-Centric Perception (AAAI 2024 · similarity 0.85)
- A Pedestrian is Worth One Prompt: Towards Language Guidance Person Re-Identification (CVPR 2024 · similarity 0.85)
- Understanding Model Reprogramming for CLIP via Decoupling Visual Prompts (ICML 2025 · similarity 0.84)
- MDReID: Modality-Decoupled Learning for Any-to-Any Multi-Modal Object Re-Identification (NeurIPS 2025 · similarity 0.84)
- Multi-Modal Object Re-identification via Sparse Mixture-of-Experts (ICML 2025 · similarity 0.84)
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.