Research map / Label-efficient and robust learning / Domain adaptation and generalization
Day-Night Cross-domain 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 DNDM, a ResNet-50 vehicle re-ID framework with glare suppression, dual-domain structure enhancement, and cross-domain class awareness modules for day-night cross-domain vehicle ReID, plus two new benchmarks (DN-Wild, DN-348).
Problem
Vehicle re-identification across day and night images is largely neglected; nighttime images suffer from headlight glare, low light, and day-night domain gaps.
Approach
Combines a glare-prompt mask-based suppression module on nighttime features, gradient-based structural descriptors added to appearance features, and class-activation-map exchange between day and night samples, inserted at multiple ResNet-50 stages to modulate day-night features.
Key results
- On DN-348 day-to-night, DNDM reaches 70.7% Rank-1 and 47.5% mAP vs 67.4% and 45.6% for the baseline.
- On DN-Wild day-to-night, DNDM reaches 51.2% Rank-1 and 40.5% mAP vs BOT's 50.3% and 39.9%, with gains limited on night-to-day (49.5% Rank-1 vs 48.5%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DNDM instead of BOT | vehicle ReID · DN-348 Day-to-Night | Rank-1: 0.663 -> 0.707 | Helped |
| use DNDM instead of BOT | vehicle ReID · DN-348 Night-to-Day | mAP: 0.453 -> 0.462 | Helped |
| use DNDM instead of BOT | vehicle ReID · DN-Wild Day-to-Night | mAP: 0.399 -> 0.405 | Helped |
| use DNDM instead of BOT | vehicle ReID · DN-Wild Night-to-Day | mAP: 0.398 -> 0.400 | Helped |
| add NGS only (ablation) | ResNet-50 vehicle ReID · DN-348 Day-to-Night | Rank-1: 0.674 -> 0.690 | Helped |
| add DSE only (ablation) | ResNet-50 vehicle ReID · DN-348 Day-to-Night | Rank-1: 0.674 -> 0.681 | Helped |
| add CCA only (ablation) | ResNet-50 vehicle ReID · DN-348 Day-to-Night | Rank-1: 0.674 -> 0.684 | Helped |
| add NGS and DSE (ablation) | ResNet-50 vehicle ReID · DN-348 Night-to-Day | Rank-1: 0.723 -> 0.800 | Helped |
| add NGS and DSE (ablation) | ResNet-50 vehicle ReID · DN-348 Day-to-Night | mAP: 0.456 -> 0.469 | Helped |
| add DSE and CCA alongside NGS (ablation) | ResNet-50 vehicle ReID · DN-348 Day-to-Night | Rank-1: 0.699 -> 0.707 | Helped |
| add DSE and CCA alongside NGS (ablation) | ResNet-50 vehicle ReID · DN-348 Night-to-Day | mAP: 0.460 -> 0.462 | Helped |
Surprising findings
- State-of-the-art visible-infrared person ReID methods gave little improvement over the BOT strong baseline on day-night vehicle data.
- PMT performs well on DN-348 but not on DN-Wild, which the authors attribute to severe sample imbalance.
Limitations
- Relies on a binary glare mask from a fixed brightness threshold of 220.
- DN-Wild shows strong day-night sample imbalance; method gains on night-to-day are small.
- Evaluated only on two datasets derived from vehicle surveillance data.
Experimental setup
Datasets: DN-Wild DN-348
Benchmarks: DN-348 DN-Wild
Metrics: Rank-1 Rank-5 mAP
Baselines: BOT DDAG LbA AGW CAJ DCL PMT Baseline
Builds on: BOT class activation mapping visual prompts
Closest papers by meaning
- NightReID: A Large-Scale Nighttime Person Re-Identification Benchmark (AAAI 2025 · similarity 0.85)
- NightAdapter: Learning a Frequency Adapter for Generalizable Night-time Scene Segmentation (CVPR 2025 · similarity 0.81)
- DNIT: Enhancing Day-Night Image-to-Image Translation through Fine-Grained Feature Handling (Student Abstract) (AAAI 2024 · similarity 0.81)
- Boosting Object Detection with Zero-Shot Day-Night Domain Adaptation (CVPR 2024 · similarity 0.81)
- Exploring the Common Appearance-Boundary Adaptation for Nighttime Optical Flow (ICLR 2024 · similarity 0.81)
- NightCC: Nighttime Color Constancy via Adaptive Channel Masking (CVPR 2024 · similarity 0.81)
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.