Research map / Video / Video segmentation and tracking
NightRain: Nighttime Video Deraining via Adaptive-Rain-Removal and Adaptive-Correction
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 NightRain, a transformer-based conditional video diffusion model for nighttime video deraining that is adapted on unlabeled real rain and clear night videos via confidence-based teacher-student self-training and difference-based correction.
Problem
Nighttime video deraining methods trained on synthetic data suffer a domain gap with real rain under complex light effects and low light, causing residual rain streaks, over-saturation, and color shifts.
Approach
Adapt a synthetic-pretrained video diffusion model to real nighttime videos without ground truth: (1) adaptive-rain-removal uses variance across multiple noise-sampled predictions as a confidence map, keeps high-confidence regions as pseudo-pairs for a student trained with noise/mask augmentation, and updates the teacher by EMA; (2) adaptive-correction trains the student on pairs formed from clear
Key results
- On SynNightRain, NightRain reaches 26.73 dB PSNR and 0.8647 SSIM, versus 22.21 dB / 0.6723 for MetaRain (the strongest baseline in the table is FDM at 23.49 dB / 0.7657).
- Ablation on real videos: Hyper-IQA scores rise from 0.6401 (pre-trained) to 0.6540 with adaptive-correction and 0.6791 with both adaptive-correction and adaptive-rain-removal.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NightRain instead of FDM | nighttime video deraining on SynNightRain · SynNightRain test | PSNR: 23.49 -> 26.73 | Helped |
| Use NightRain instead of FDM | nighttime video deraining on SynNightRain · SynNightRain test | SSIM: 0.7657 -> 0.8647 | Helped |
| Add adaptive-correction to the pretrained model (ablation) | NightRain · RealNightRain | Hyper-IQA: 0.6401 -> 0.6540 | Helped |
| Add adaptive-rain-removal to adaptive-correction (ablation) | NightRain · RealNightRain | Hyper-IQA: 0.6540 -> 0.6791 | Helped |
| Add adaptive-correction to the pretrained model (ablation) | NightRain · RealNightRain | color fidelity | Helped |
| Add adaptive-rain-removal to adaptive-correction (ablation) | NightRain · RealNightRain | rain streak removal | Helped |
Limitations
- Relies on synthetic pretraining on SynNightRain to initialize the diffusion model.
- Real-world evaluation is qualitative plus a non-reference metric; no paired real ground truth.
- Dark high-confidence regions (e.g., sky) can be retained by the confidence criterion, which only addresses deraining.
Experimental setup
Datasets: SynNightRain RealNightRain (unlabeled real rain videos) real clear nighttime videos
Benchmarks: SynNightRain RealNightRain
Metrics: PSNR SSIM Hyper-IQA (non-reference)
Baselines: HRIR DLF RMFD DSTFM MetaRain WeatherDiff FDM
Builds on: denoising diffusion probabilistic models DiT (adaLN transformer blocks) mean teacher (EMA)
Closest papers by meaning
- Rethinking Nighttime Image Deraining via Learnable Color Space Transformation (NeurIPS 2025 · similarity 0.91)
- NightHaze: Nighttime Image Dehazing via Self-Prior Learning (AAAI 2025 · similarity 0.86)
- Genuine Knowledge from Practice: Diffusion Test-Time Adaptation for Video Adverse Weather Removal (CVPR 2024 · similarity 0.86)
- Harnessing Joint Rain-/Detail-aware Representations to Eliminate Intricate Rains (ICLR 2024 · similarity 0.86)
- NightCC: Nighttime Color Constancy via Adaptive Channel Masking (CVPR 2024 · similarity 0.86)
- Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining (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.