Research map / Label-efficient and robust learning / Domain adaptation and generalization
NightCC: Nighttime Color Constancy via Adaptive Channel Masking
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 an unsupervised domain adaptation mean-teacher framework for nighttime color constancy that uses adaptive channel masking and pixel-wise light uncertainty to learn from unlabeled night images with labeled daytime data.
Problem
Color constancy (illuminant estimation) under nighttime conditions with multiple colored light sources and low-light sensor noise, where models trained on daytime/indoor data generalize poorly and pixel-level nighttime labels are costly.
Approach
Combines mean-teacher semi-supervised learning with (1) adaptive channel masking, which identifies feature channels whose covariance is insensitive to Night Augmentation (simulated multi-light colors and sensor read noise) and masks them in the student, and (2) light uncertainty, a pixel-wise confidence map from a two-pass teacher prediction used to weight the unsupervised loss.
Key results
- Achieves mean MAE 4.82 and median 4.69 on the nighttime dataset, versus 6.19 for the best prior multi-illuminant method Gray-Index and 6.61 for CLCC (a 23.7% reduction vs CLCC in mean MAE).
- Ablation: mean MAE falls from 7.36 (daytime only) to 6.95 with unlabeled night data, 5.40 with light uncertainty, and 4.82 with adaptive channel masking added.
- On LSMI with 10% labeled data, mean MAE is 3.04 with 90% unlabeled data versus 5.71 with no unlabeled data (33.8% improvement).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NightCC instead of MIMT | Nighttime color constancy · Nighttime dataset | mean MAE: 6.14 -> 4.82 | Helped |
| Use NightCC instead of MIMT | Nighttime color constancy · Nighttime dataset | median MAE: 5.84 -> 4.69 | Helped |
| Use NightCC with 10% labeled and 90% unlabeled training data instead of MIMT | Color constancy · LSMI test | mean MAE: 2.48 -> 3.04 | Hurt |
| Add unlabeled nighttime data (ablation) | NightCC · Nighttime dataset | mean MAE: 7.36 -> 6.95 | Helped |
| Add light uncertainty (ablation) | NightCC with unlabeled nighttime data · Nighttime dataset | mean MAE: 6.95 -> 5.40 | Helped |
| Add adaptive channel masking (ablation) | NightCC with light uncertainty · Nighttime dataset | mean MAE: 5.40 -> 4.82 | Helped |
| Add unlabeled nighttime data (ablation) | NightCC · Nighttime dataset | median MAE: 6.65 -> 6.21 | Helped |
| Add light uncertainty (ablation) | NightCC with unlabeled nighttime data · Nighttime dataset | median MAE: 6.21 -> 4.97 | Helped |
| Add adaptive channel masking (ablation) | NightCC with light uncertainty · Nighttime dataset | median MAE: 4.97 -> 4.69 | Helped |
| Use unlabeled training data in addition to 10% labeled data (ablation) | NightCC · LSMI test | mean MAE: 5.71 -> 3.04 | Helped |
Surprising findings
- On LSMI, the semi-supervised model with only 10% labels (5.71) does not beat fully supervised Patch CNN (4.82) or AngularGAN (4.69), and MIMT (2.48) substantially outperforms it at full supervision.
Limitations
- Nighttime dataset is small (64 labeled evaluation images) and relies on sparse annotations from achromatic paper.
- Semi-supervised results on LSMI with limited labels lag behind some fully supervised baselines.
- Training uses fixed 256x256 resolution and NVIDIA RTX 3090 GPUs; code and data release not stated.
Experimental setup
Datasets: Sony daytime dataset NUS-8 LSMI nighttime dataset (720 unlabeled, 64 labeled)
Benchmarks: Nighttime dataset (own) LSMI
Metrics: mean angular error median angular error
Baselines: CLCC Quasi-U AngularGAN MIMT MSCC FFCC FC4 Gray-Pixel Gray-Index White-Patch
Builds on: mean teacher U-Net AngularGAN Quasi-U
Closest papers by meaning
- Rethinking Nighttime Image Deraining via Learnable Color Space Transformation (NeurIPS 2025 · similarity 0.88)
- See through the Dark: Learning Illumination-affined Representations for Nighttime Occupancy Prediction (NeurIPS 2025 · similarity 0.87)
- Exploring the Common Appearance-Boundary Adaptation for Nighttime Optical Flow (ICLR 2024 · similarity 0.87)
- NightRain: Nighttime Video Deraining via Adaptive-Rain-Removal and Adaptive-Correction (AAAI 2024 · similarity 0.86)
- Learning Adaptive Lighting via Channel-Aware Guidance (ICML 2025 · similarity 0.85)
- NightHaze: Nighttime Image Dehazing via Self-Prior Learning (AAAI 2025 · similarity 0.85)
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.