Research map / Label-efficient and robust learning / Domain adaptation and generalization
NightAdapter: Learning a Frequency Adapter for Generalizable Night-time Scene Segmentation
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 NightAdapter, a parameter-efficient frequency-domain side adapter on a frozen DINOv2 that adapts illumination-insensitive low-frequency bands and randomizes-then-adapts illumination-sensitive bands via discrete sine transform, for generalizable night-time segmentation.
Problem
Night-time semantic segmentation models generalize poorly to unseen night scenes with varying illumination, both when trained on day-time data (day-to-night) and when trained on one night dataset (night-to-night).
Approach
Uses the discrete sine transform on frozen VFM features to split frequency bands into illumination-insensitive (low-frequency) and illumination-sensitive (middle/high) groups; adapts the former with a token-based side adapter and randomizes then adapts the latter, bringing the frequency-prior idea of domain generalization into a PEFT scheme for VFMs.
Key results
- Day-to-night (Cityscapes source): 84.0 mIoU on ACDC-Night vs 70.6 for REIN baseline (+13.4), and 74.1 on Dark Zurich (+9.8).
- In-domain night segmentation: 65.1 mIoU on NightCity, 66.2 on NightCity-fine, 57.8 on BDD-Night, outperforming REIN by 2.7, 1.6, 3.9 points with 11.5M trainable parameters.
- Night-to-night generalization with NightCity source: 50.5 on ACDC-Night, improving over the best prior VFM method by up to 5.2 mIoU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add IBA to the REIN baseline (ablation) | NightAdapter · AN, DZ, ND | mIoU: +1.9%, +3.5%, +2.4% | Helped |
| Add SBR (randomize illumination-sensitive bands) without SBA (ablation) | NightAdapter · AN, DZ, ND | mIoU: +3.5%, +2.6%, +2.1% | Helped |
| Add SBA to the configuration with IBA and SBR (ablation) | NightAdapter · AN, DZ, ND | mIoU: +8.0%, +3.7%, +5.5% | Helped |
| Adapt additional frequency bands, adding VM1, then VM2, then VH alongside VL (ablation) | NightAdapter · AN, DZ, ND | mIoU: AN: 72.5, 74.8, 77.2, 84.0; DZ: 67.8, 69.7, 71.5, 74.1; ND: 52.7, 54.6, 56.7, 60.3 | Helped |
| Use the full NightAdapter instead of REIN | NightAdapter · AN, DZ, ND (day-to-night) | mIoU: +13.4, +9.8, +10.0 | Helped |
| Use the full NightAdapter instead of REIN | NightAdapter · NC, NC-f, BN | mIoU: +2.7%, +1.6%, +3.9% | Helped |
| Use NightAdapter instead of REIN with NC as source | NightAdapter · AN, DZ, ND, BN, AC | mIoU: +5.2, +3.0, +1.1, +5.7, +3.7 | Helped |
| Use NightAdapter instead of REIN with NC-f as source | NightAdapter · AN, DZ, ND, BN, AC | mIoU: +4.1%, +3.0%, +1.5%, +3.2%, +3.1% | Helped |
Surprising findings
- Removing low-frequency bands hurts generalization on both source and target domains, while removing high-frequency bands improves it.
Limitations
- Relies on the REIN training recipe and hyperparameters, with a fixed randomization threshold T=0.3 chosen empirically.
- Gains on Nighttime Driving (ND) are modest compared with other benchmarks (+1.1 over best night-to-night baseline).
Experimental setup
Datasets: Cityscapes NightCity NightCity-fine BDD100K-Night
Benchmarks: ACDC-Night Dark Zurich Nighttime Driving BDD100K-Night All-day CityScapes NightCity NightCity-fine
Metrics: mIoU
Baselines: REIN DAFormer HRDA MIC CoDA PASS SET FADA tqdm LoRA AdaptFormer VPT
Builds on: REIN DINOv2 Mask2Former LoRA
Closest papers by meaning
- The Parables of the Mustard Seed and the Yeast: Extremely Low-Budget, High-Performance Nighttime Semantic Segmentation (AAAI 2025 · similarity 0.87)
- Exploring the Common Appearance-Boundary Adaptation for Nighttime Optical Flow (ICLR 2024 · similarity 0.85)
- NightCC: Nighttime Color Constancy via Adaptive Channel Masking (CVPR 2024 · similarity 0.84)
- DNIT: Enhancing Day-Night Image-to-Image Translation through Fine-Grained Feature Handling (Student Abstract) (AAAI 2024 · similarity 0.84)
- NightRain: Nighttime Video Deraining via Adaptive-Rain-Removal and Adaptive-Correction (AAAI 2024 · similarity 0.83)
- Rethinking Nighttime Image Deraining via Learnable Color Space Transformation (NeurIPS 2025 · similarity 0.83)
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.