Research map / Label-efficient and robust learning / Recognition and long-tail
NightReID: A Large-Scale Nighttime Person Re-Identification Benchmark
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 NightReID, a 53k-image RGB nighttime person re-ID dataset, and the EDA framework (unsupervised enhancement/denoising IED plus statistical data distribution alignment DDA) that improves TransReID-SSL on nighttime benchmarks.
Problem
Person re-identification under nighttime RGB conditions is hard due to low light, noise, and distribution shift from daytime pre-training data, and existing nighttime datasets are either infrared/cross-modal or dominated by extremely dark images.
Approach
Combines a no-reference Zero-DCE-style enhancement network with a detail-extraction denoising step and a loss that adapts enhancement to brightness (IED), plus z-score normalization of the dataset with its own channel statistics to align nighttime data with pre-training distributions (DDA), running them in parallel branches.
Key results
- EDA reaches 67.6 Rank-1 and 52.3 mAP on NightReID (528 IDs with distractors), up from 65.0/50.8 for the TransReID-SSL baseline.
- On Night600, EDA reaches 23.8 Rank-1 and 11.6 mAP, versus 20.0/9.5 for the baseline and 19.2/9.5 for CENet.
- Ablation: DDA alone on the baseline improves Night600 Rank-1 from 20.0 to 22.1; combining IED and DDA in separate branches gives the best results on both datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add EDA to TransReID-SSL | TransReID-SSL · NightReID (528 IDs w/ distractors) | Rank-1: 65.0 -> 67.6 | Helped |
| Add EDA to TransReID-SSL | TransReID-SSL · Night600 | Rank-1: 20.0 -> 23.8 | Helped |
| Add DDA to the baseline (ablation) | TransReID-SSL · NightReID | Rank-1: 65.0 -> 66.1 | Helped |
| Add DDA to the baseline (ablation) | TransReID-SSL · Night600 | Rank-1: 20.0 -> 22.1 | Helped |
| Replace vanilla Zero-DCE with IED in a single branch (ablation) | TransReID-SSL · NightReID | Rank-1: 62.9 -> 67.0 | Helped |
| Replace vanilla Zero-DCE with IED in a single branch (ablation) | TransReID-SSL · Night600 | Rank-1: 19.1 -> 20.2 | Helped |
| Add IED in a single branch (ablation) | TransReID-SSL · NightReID | Rank-1: 65.0 -> 67.0 | Helped |
| Add IED in a single branch (ablation) | TransReID-SSL · Night600 | Rank-1: 20.0 -> 20.2 | No clear effect |
| Combine IED and DDA in separate branches (ablation) | TransReID-SSL · NightReID | Rank-1: 65.0 -> 67.6 | Helped |
| Combine IED and DDA in separate branches (ablation) | TransReID-SSL · Night600 | Rank-1: 20.0 -> 23.8 | Helped |
Surprising findings
- Vanilla Zero-DCE enhancement lowers NightReID performance versus the baseline (62.9 vs 65.0 Rank-1), indicating that enhancement can hurt Re-ID.
- PFD-Net and DC-Former perform worse than their baselines on the nighttime benchmarks.
Limitations
- Night600 images are extremely dark, causing convergence issues and poor performance for many methods.
- IED and DDA use separate branches and features are concatenated at test time, which adds architectural complexity; a single module in a single branch gives weaker gains.
Experimental setup
Datasets: NightReID Night600 LUPerson
Benchmarks: NightReID Night600
Metrics: Rank-1 mAP CMC
Baselines: BoT AGW MSINet HAT TransReID TransReID-SSL PFD-Net DC-Former CLIP-ReID IDF CENet
Builds on: Zero-DCE TransReID-SSL BoT CLAHE
Closest papers by meaning
- Day-Night Cross-domain Vehicle Re-identification (CVPR 2024 · similarity 0.85)
- Feature-Level Adversarial Attacks and Ranking Disruption for Visible-Infrared Person Re-identification (NeurIPS 2024 · similarity 0.85)
- AG-VPReID: A Challenging Large-Scale Benchmark for Aerial-Ground Video-based Person Re-Identification (CVPR 2025 · similarity 0.85)
- Alice Benchmarks: Connecting Real World Re-Identification with the Synthetic (ICLR 2024 · similarity 0.85)
- From Laboratory to Real World: A New Benchmark Towards Privacy-Preserved Visible-Infrared Person Re-Identification (CVPR 2025 · similarity 0.84)
- Temporal Correlation Vision Transformer for Video Person Re-Identification (AAAI 2024 · 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.