Research map / Detection and segmentation / Open-set and few-shot detection
The Parables of the Mustard Seed and the Yeast: Extremely Low-Budget, High-Performance Nighttime Semantic 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.
Introduces active domain adaptive nighttime semantic segmentation (ADA-NSS) with a region-based query strategy (CARQ) and a fragmentation-synergy framework (FS-ADA) that propagates sparse labels, reaching fully supervised-level mIoU with 1% target annotation.
Problem
Nighttime semantic segmentation relies on massive pixel-level labels; existing active domain adaptation methods for segmentation are designed for day/synthetic domains and fail under the large day-to-night domain gap with very low labeling budgets.
Approach
Query image patches using a Value Mixture of Experts combining a distraction-perception score (weather-specific degradation via frequency components and saliency) with a semantic-complexity score, then propagate the few labels to neighboring and global regions via prototype-based label propagation and consistency losses.
Key results
- On BDD100K-Night val, FS-ADA with 1% target annotation reaches 42.0 mIoU with HRDA? (HRDA-based FS-ADA exceeds fully supervised HRDA on ACDC-night-val: 66.3 vs 67.4 on ACDC-night-test per Table 1; DAFormer-based 59.1 vs 62.0)
- With HRDA backbone, FS-ADA (1%) reaches 66.3 mIoU on ACDC-night-test, versus 63.5 for Refign and 65.1 for InforMS.
- Ablation on ACDC-night-val with DAFormer: warmup restoration and segmentation raise mIoU from 45.9 to 49.2; adding CARQ, LPM and LGM brings it to 52.1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FS-ADA instead of D2ADA with DeepLab-v2 | ADA-NSS on DeepLab-v2 · ACDC-night-test | mIoU: 52.0 -> 52.4 | Helped |
| Use FS-ADA instead of D2ADA with DeepLabv3+ | ADA-NSS on DeepLabv3+ · ACDC-night-test | mIoU: 50.4 -> 52.4 | Helped |
| Use FS-ADA instead of fully supervised DeepLab-v2 | DeepLab-v2 segmentation · BDD100K-Night val | mIoU: 36.3 -> 37.7 | Helped |
| Fine-tune the restoration network to target data (ablation) | Warmup framework with DAFormer · ACDC-night-val | mIoU: 45.9 -> 47.0 | Helped |
| Constrain prediction consistency with aligned target fragments (ablation) | Warmup framework with DAFormer · ACDC-night-val | mIoU: 47.0 -> 48.0 | Helped |
| Optimize complementary-class learning (ablation) | Warmup framework with DAFormer · ACDC-night-val | mIoU: 48.0 -> 49.2 | Helped |
| Fine-tune the restoration network to target data (ablation) | Warmup framework with HRDA · ACDC-night-val | mIoU: 53.3 -> 54.7 | Helped |
| Constrain prediction consistency with aligned target fragments (ablation) | Warmup framework with HRDA · ACDC-night-val | mIoU: 54.7 -> 55.5 | Helped |
| Optimize complementary-class learning (ablation) | Warmup framework with HRDA · ACDC-night-val | mIoU: 55.5 -> 55.7 | Helped |
| Add CARQ region selection (ablation) | FS-ADA with DAFormer · ACDC-night-val | mIoU: 49.2 -> 51.0 | Helped |
| Add CARQ, label propagation, and label generalization (ablation) | FS-ADA with DAFormer · ACDC-night-val | mIoU: 49.2 -> 52.1 | Helped |
| Add CARQ region selection (ablation) | FS-ADA with HRDA · ACDC-night-val | mIoU: 55.7 -> 56.4 | Helped |
Surprising findings
- FS-ADA with 1% target annotation exceeds fully supervised DAFormer and HRDA on ACDC-night-val (+0.9 and +1.9 mIoU), per the authors' comparison.
Limitations
- Relies on a warmup UDA stage and a restoration network that must be fine-tuned, adding pipeline complexity
- Results reported in Table 2 show FS-ADA does not beat fully supervised DeepLab-v2 on several datasets (e.g., D4 BDD100K-Night 37.7 vs 36.3 is better, but D3 with DeepLab-v2 is 50.2 vs 51.2)
Experimental setup
Datasets: Cityscapes ACDC NightCity NightCity-fine BDD100K
Benchmarks: ACDC-night-test ACDC-night-val NightCity val NightCity-fine val BDD100K-Night val ACDC fog/rain/snow test
Metrics: mIoU
Baselines: DeepLab-v2 DANNet Bi-Mix Refign RIPU D2ADA HALO IA-Seg DTPS InforMS PIG VBLC
Builds on: DAFormer HRDA RIPU D2ADA Refign
Closest papers by meaning
- NightAdapter: Learning a Frequency Adapter for Generalizable Night-time Scene Segmentation (CVPR 2025 · similarity 0.87)
- See through the Dark: Learning Illumination-affined Representations for Nighttime Occupancy Prediction (NeurIPS 2025 · similarity 0.85)
- Parsing All Adverse Scenes: Severity-Aware Semantic Segmentation with Mask-Enhanced Cross-Domain Consistency (AAAI 2024 · similarity 0.85)
- NightRain: Nighttime Video Deraining via Adaptive-Rain-Removal and Adaptive-Correction (AAAI 2024 · similarity 0.85)
- Exploring the Common Appearance-Boundary Adaptation for Nighttime Optical Flow (ICLR 2024 · similarity 0.84)
- Weakly Supervised Semantic Segmentation for Driving Scenes (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.