In computer vision, reconstructing the 3D shape of an object from 2D images is often thwarted by a fundamental ambiguity: is a dark patch on a surface a shadow caused by geometry, or is it simply a dark mark, like black ink or a deep skin tone? This confusion leads to "texture-copying" artifacts, where high-contrast surface patterns are mistakenly "baked into" the recovered 3D model as physical indentations or bumps.
To solve this, researchers have developed a family of techniques that decouple surface reflectance (albedo) from geometric shading. The core intuition is that while albedo is an intrinsic property of the material, shading is a function of the light's direction. By comparing multiple images of the same point under different conditions, we can mathematically cancel out the albedo, leaving only the pure geometric signal.
The Algebraic Cancellation of Albedo
The most direct way to eliminate surface color is through the image ratio. In a standard Lambertian reflection model, the intensity of a pixel is the product of the surface albedo ($\rho$) and the cosine of the angle between the surface normal ($n$) and the light source ($l$).
$$I = \rho (n \cdot l)$$
If we take two images ($I_1$ and $I_2$) of the same point under two different light sources ($l_1$ and $l_2$), the albedo $\rho$ remains constant. By simply dividing one image by the other, the albedo term cancels out entirely:
$$\frac{I_1}{I_2} = \frac{\rho (n \cdot l_1)}{\rho (n \cdot l_2)} = \frac{n \cdot l_1}{n \cdot l_2}$$
As Yvain Quéau, Roberto Mecca, and Jean-Denis Durou demonstrate in Unbiased Photometric Stereo for Colored Surfaces: A Variational Approach, this ratio transforms a non-linear shading problem into a system of linear partial differential equations. This allows for shape reconstruction that is completely independent of surface color patterns.
This principle is remarkably versatile. Hideo Saito and Seong Ik Cho applied this to the problem of reconstructing the curved surfaces of open books. By isolating the "ink" from the "light," they prevented text from being misinterpreted as geometric depth. Similarly, Xu Cao and Boxin Shi showed that the ratio between flash and no-flash images can isolate surface normals, preventing dark textures from appearing as physical holes in the 3D mesh.
Logarithmic Invariance and Event Cameras
A more modern take on this concept involves event-based sensors. Unlike traditional cameras that capture frames, event cameras only report changes in brightness at each pixel.
Because these sensors respond to the logarithm of intensity changes, the albedo naturally drops out of the equation. If $I = \rho S$ (where $S$ is shading), then $\log(I) = \log(\rho) + \log(S)$. When the sensor calculates a temporal difference ($\frac{d}{dt}\log I$), the static albedo term $\log(\rho)$ disappears.
Bohan Yu and Jin Han leverage this in EventPS and EventPSR. By using a rotating light source, they can estimate high-resolution surface normals in real-time. The sensing mechanism itself acts as a hardware-level albedo filter, making the system inherently robust to complex surface patterns without requiring additional color-correction software.
Photometric Invariants and Blind Elimination
Beyond simple ratios, researchers have sought "photometric invariants"—mathematical representations that remain unchanged regardless of the object's material or the lighting direction.
In their seminal work, Visvanathan Ramesh and Shree Nayar introduced a framework for separating material from shape using determinant ratios of image matrices. By treating pixel measurements as an algebraic structure, they could "blindly" eliminate unknown variables like light source intensity or surface albedo. This allows for the creation of "material maps" that are entirely flat, showing the true color of an object without any shadows or highlights.
This concept of invariance extends to complex neural representations as well. When aligning two different Neural Radiance Fields (NeRFs) captured at different times, Daniel Rebain and Sara Sabour distilled a lighting-invariant surface field. By focusing on the underlying density (geometry) rather than the view-dependent radiance (color), they enabled precise 3D registration even when the lighting conditions were completely different between the two captures.
Why It Works: The Geometric Anchor
The reason these ratio-based and invariant methods are so successful is that they treat geometry as the anchor. While lighting and albedo can vary wildly, the physical orientation of a surface point is a constant.
Whether it is Edwin Hancock using intensity ratios to estimate surface roughness or Boxin Shi using albedo ratios to calibrate millions of internet photos, the strategy is the same: find a mathematical transformation where the material properties cancel out, leaving the geometry as the only variable left to explain the data. By doing so, computer vision moves from merely "seeing" pixels to "understanding" the physical structures that created them.