Deep learning models excel at visual tasks when provided with massive, pixel-perfect datasets. However, when deployed in low-light environments, dense fog, or heavy rain, their performance plummets. The intuitive solution would be to gather and manually label thousands of nighttime or adverse-weather images. But this approach hits a fundamental bottleneck: humans cannot reliably label what they cannot see.
As noted by Van Gool et al. (2018) and Lee et al. (2023), manual pixel-level annotation of nighttime imagery is exceptionally difficult and error-prone due to underexposure, sensor noise, and severe glare from artificial lights. Object boundaries become obscured, making it nearly impossible for human annotators to accurately delineate shapes or identify keypoints (Ai et al. (2024)). This creates a severe data scarcity that prevents traditional supervised learning from scaling to 24/7 operations (Li et al. (2021)).
To overcome this, a rapidly growing body of research focuses on unsupervised domain adaptation (UDA) and synthetic generation—methods designed to teach models to see in the dark without requiring a single manual low-light annotation.
Leveraging Invariant Structural Anchors
The core intuition behind bypassing the low-light bottleneck is simple: while illumination changes drastically between day and night, the underlying geometry and structure of the physical world remain identical. A road, a building, or a traffic sign does not move when the sun goes down.
Researchers exploit this spatial invariance by using cross-modal or cross-temporal anchors to transfer knowledge from the well-lit daytime domain to the dark nighttime domain. For instance, Lee et al. (2023) introduced GPS-GLASS, a framework that uses GPS data to precisely align daytime video frames with nighttime scenes. Because the frames are spatially matched, the high-quality predictions or labels from the daytime images can be directly used as pseudo-labels to train the network on nighttime imagery. Similarly, Sakaridis (2023) developed Refign, which utilizes GPS-matched reference images from clear weather to align and fuse daytime predictions into the training process for adverse conditions like night, fog, or snow.
Beyond spatial alignment, other invariant properties can act as structural "cheat codes." Wang et al. (2025) leverage depth maps within a diffusion-guided knowledge distillation framework, noting that geometric depth remains consistent regardless of lighting variations. When vision completely fails, alternative modalities can step in: Liu et al. (2025) utilize an adaptive audio-visual fusion mechanism that naturally leans on sound cues to maintain activity recognition accuracy when the visual domain degrades.
Synthesizing the Dark via Physics and Generative Models
If collecting and labeling real low-light data is too difficult, another powerful strategy is to synthetically darken perfectly labeled daytime datasets. However, simply lowering the brightness of an image does not capture the complex physics of nighttime environments, which include sensor noise, headlight glare, and distinct material reflections.
To bridge this realism gap, researchers use physics-based decomposition and simulation. Du et al. (2024) employ Retinex theory to decouple an image into its structural reflectance and its illumination. Because reflectance represents the intrinsic properties of objects, the model can be trained to remain invariant to lighting changes, achieving zero-shot day-night adaptation. Taking a generative approach, Van Gool (2025) developed SOLO, a pipeline that reconstructs a single daytime image in 3D to explicitly model geometry, materials, and artificial light sources, generating photorealistic nighttime counterparts that inherit the original daytime labels.
Other approaches focus on simulating the specific degradation patterns of low-light sensors. Zheng et al. (2020) address the failure of optical flow algorithms in the dark by training networks on synthetic datasets that realistically model raw sensor noise. By learning to handle noise directly at the sensor level, models can estimate motion without relying on error-prone "enhance-then-estimate" pipelines.
Self-Supervision and Bootstrapping
When paired daytime data or explicit physical models are unavailable, networks must learn to bootstrap directly from unlabeled low-light data. This is typically achieved through teacher-student distillation and self-consistency loops.
In a typical setup, a model trained on daytime data acts as a teacher, generating pseudo-labels for nighttime images. To prevent the student model from learning the teacher's errors, frameworks must filter out low-quality predictions. Wu et al. (2021) introduced DANNet, a one-stage domain adaptation network that uses a re-weighting strategy to handle coarsely aligned day-night pairs. To enforce stricter structural consistency, the authors of LoopDA (2023) introduced "self-loops," a mechanism that takes the predicted semantic maps of nighttime images and renders them back into latent features to ensure the predictions remain physically plausible.
Recently, the emergence of Vision Foundation Models (VFMs) has provided a powerful new tool for bootstrapping. Yao et al. (2023) developed SAM-DA, which leverages the zero-shot segmentation power of the Segment Anything Model (SAM) to automatically generate high-quality training samples from a small set of raw nighttime images. This process, termed "sample swelling," allows tracking models to adapt to nocturnal environments while requiring 90% less raw data, effectively breaking the annotation bottleneck.
Key Foundational Work
Dark Model Adaptation: Semantic Image Segmentation from Daytime to Nighttime (2018) — contributed 1 concept elements to this direction.
Guided Curriculum Model Adaptation and Uncertainty-Aware Evaluation for Semantic Nighttime Image Segmentation (2019) — contributed 1 concept elements to this direction.
Curriculum Model Adaptation with Synthetic and Real Data for Semantic Foggy Scene Understanding (2019) — contributed 1 concept elements to this direction.