Spatially Varying Atmospheric Light for Image Restoration

This direction is split into other directions: spatially varying atmospheric light modeling. Its last version stays readable and citable here.

Version 1 of 135 papers57 conceptsfrozen 2026-10-04

In classical computer vision, the physics of how light interacts with the atmosphere is often distilled into a simple equation: the Atmospheric Scattering Model (ASM). This model describes a hazy image as a linear combination of the clear scene and a "global" atmospheric light. For years, researchers assumed this light was a single, uniform constant across the entire image.

However, real-world environments—ranging from neon-lit city streets at night to complex indoor scenes with multiple lamps—rarely obey this rule. Recent research has shifted toward modeling atmospheric light as a spatially varying field, a transition that has unlocked significant improvements in dehazing, shadow removal, and 3D reconstruction.

The Failure of the Global Constant

The traditional assumption of uniform atmospheric light is a mathematical convenience that fails in the "wild." When a model assumes the light source is a single global constant, it essentially tries to "average out" the illumination of the entire scene.

As noted by Yan Huang and colleagues in their work on FWB-Net, this simplification leads to two major artifacts:

  1. Color Distortion: In sky regions or areas near bright objects, the model over-corrects, creating unnatural color shifts or a "grey tent" effect.
  2. Halo Artifacts: Around localized light sources, the model fails to account for the glow (scattering), leading to dark halos or oversaturation.

By treating atmospheric light as a pixel-location-dependent variable, researchers like Shengdong Zhang in the Color-Constrained Dehazing Model allow the restoration process to adapt to local variations in light intensity and color. This prevents the "baking in" of lighting errors into the final restored image.

From Daytime Fog to Nighttime Glow

The limitations of global light models are most obvious at night. While daytime haze is often dominated by a single distant source (the sun), nighttime scenes are a "wild west" of artificial lighting.

Jing Zhang and Yang Cao have pioneered methods for Nighttime Haze Removal that model ambient illumination as a spatially varying component. Instead of one $A$ value for the whole image, they estimate a local light map. This allows the algorithm to distinguish between the haze itself and the multi-colored glow emanating from street lamps or neon signs. Their later work on the Maximum Reflectance Prior further refined this by focusing on the brightest pixels in local patches to estimate these varying light colors 10 to 100 times faster than previous methods.

Decoupling Light from Geometry

The intuition behind spatially varying light extends beyond just cleaning up foggy photos; it is critical for understanding 3D shapes. When reconstructing a 3D object, if a model assumes light is uniform, it might mistake a shadow for a physical hole or a bright highlight for a change in surface color (albedo).

In Intrinsic3D, Robert Maier and Matthias Nießner demonstrate that representing illumination as a spatially varying field prevents local shadows from being misinterpreted as surface geometry. This allows consumer-grade sensors to recover fine-scale details—like the texture of a fabric or the subtle curves of a face—that would otherwise be "melted" away by simpler lighting models.

Similarly, in the realm of shadow removal, Lanqing Guo and the creators of ShadowDiffusion use a spatially-variant degradation model. By decomposing a shadow into a binary mask and a continuous illumination weight, they can precisely model how light is obstructed across different regions, leading to cleaner boundaries and more realistic restoration.

Physics-Aware Neural Estimation

The most recent trend, exemplified by Huaizhuo Liu in PEIE, is to embed these physical insights directly into neural network architectures. Instead of letting a "black box" CNN guess the clear image, these networks are designed to explicitly estimate an Illumination-Adaptive Scattering Model (IASM).

By forcing the network to first predict the spatially varying light map, the model gains "physics-awareness." It understands that a pixel might be bright because it's close to a light source, not just because the haze is thick. This hybrid approach—combining the flexibility of deep learning with the constraints of spatially varying physics—represents the current state-of-the-art in making computer vision robust enough for the real world.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_spatially_varying_atmospheric_light_for__368d72fd,
  title = {Spatially Varying Atmospheric Light for Image Restoration},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spatially-varying-atmospheric-light-for-image-restoration/rdr_368d72fd6ef74e0ca1dc2952e3b3042c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_spatially_varying_atmospheric_light_for__368d72fd_v1,
  title = {Spatially Varying Atmospheric Light for Image Restoration},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spatially-varying-atmospheric-light-for-image-restoration/rdr_368d72fd6ef74e0ca1dc2952e3b3042c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}