Albedo Invariant Shape Recovery via Intensity Ratios

61 concepts44 papers1 proposal

When we look at an image, the brightness of any given pixel is determined by two primary factors: the shape of the object (how its surface tilts relative to the light) and the material of the object (its intrinsic color or texture, known as albedo). For decades, computer vision algorithms like shape-from-shading and photometric stereo have tried to untangle these two properties to reconstruct 3D surfaces from 2D images.

However, as Micah K. Johnson et al. point out in Retrographic sensing for the measurement of surface texture and shape, standard shading-based methods fundamentally struggle to distinguish between a change in surface color and an actual geometric change. This ambiguity leads to "texture-copying artifacts," where high-contrast surface markings—such as text printed on a book page—are mistakenly reconstructed as physical indentations or ridges (Hideo Saito et al., A Divide-and-Conquer Strategy in Recovering Shape of Book Surface from Shading).

To solve this, a powerful paradigm has emerged: albedo-invariant shape recovery. Instead of trying to estimate the complex, spatially varying textures of an object first, these methods eliminate the albedo term entirely using mathematical invariants—most notably, intensity ratios.

The Intuition: Why the Ratio Cancels Albedo

To understand why this works, consider the simplest physics model for image formation: Lambertian reflectance. The intensity $I$ of a pixel is the product of the material's albedo $\rho$, the light source intensity $L$, and the cosine of the angle between the surface normal $N$ and the light direction $S$:

$$I = \rho \cdot L \cdot (N \cdot S)$$

If we only have one image, we cannot separate $\rho$ from $N$. A dark pixel could mean a dark material ($\text{low } \rho$) or a flat surface tilted away from the light ($\text{low } N \cdot S$).

But what happens if we capture a second image of the exact same surface point under a different light source direction, $S_2$?

$$I_1 = \rho \cdot L_1 \cdot (N \cdot S_1)$$ $$I_2 = \rho \cdot L_2 \cdot (N \cdot S_2)$$

If we divide the two equations, the unknown albedo $\rho$ (along with any uniform camera scaling) appears in both the numerator and the denominator, canceling out completely:

$$\frac{I_1}{I_2} = \frac{\rho \cdot L_1 \cdot (N \cdot S_1)}{\rho \cdot L_2 \cdot (N \cdot S_2)} = \frac{L_1 \cdot (N \cdot S_1)}{L_2 \cdot (N \cdot S_2)}$$

As David Jacobs et al. demonstrated in Comparing Images under Variable Illumination, this ratio isolates a pure geometric signal that depends exclusively on the 3D shape and the change in illumination. By taking the ratio, the algorithm becomes completely blind to the surface's visual patterns, allowing it to recover the underlying geometry without bias.

This same cancellation principle applies when estimating environmental lighting rather than shape. For instance, Imari Sato et al. (Illumination Distribution from Shadows) isolate incident light distributions by comparing shadowed and fully illuminated states of a surface, using a reference frame to cancel out the unknown material properties of the ground. Similarly, Xinxin Zuo et al. (Detailed Surface Geometry and Albedo Recovery from RGB-D Video under Natural Illumination) calculate intensity ratios between aligned video frames of a moving object to cancel albedo, enabling them to robustly solve for complex global illumination parameters like spherical harmonics.

From Ratios to Differential Equations

Once albedo is eliminated via division, the remaining challenge is to reconstruct the continuous 3D surface from these pixel-wise ratios.

Yvain Quéau et al. (Unbiased Photometric Stereo for Colored Surfaces: A Variational Approach) showed that by cross-multiplying the ratio equation ($I_2 (N \cdot S_1) - I_1 (N \cdot S_2) = 0$), the problem can be elegantly formulated as a system of linear Partial Differential Equations (PDEs). Because this system solves for the surface normals directly from the raw image intensities, it avoids the "sequential bias" of older methods that tried to estimate albedo first and geometry second.

This PDE formulation is highly adaptable. Fotios Logothetis et al. extended this approach to handle complex near-field lighting environments where light sources are close to the object (Semi-Calibrated Near Field Photometric Stereo). By embedding these intensity ratios into a volumetric Signed Distance Field (SDF), the authors transformed non-linear shading constraints into stable, linear PDEs, allowing for highly accurate reconstructions of multi-colored objects (A Differential Volumetric Approach to Multi-View Photometric Stereo).

Furthermore, this ratio constraint can be paired with other physical modalities. Silvia Tozza et al. (Linear Differential Constraints for Photo-Polarimetric Height Estimation) combined multi-light intensity ratios with polarization cues, creating a unified framework that solves directly for surface height while remaining completely invariant to spatial texture variations.

Hardware-Driven Invariance: Flash and Event Cameras

While traditional photometric stereo requires a controlled laboratory setup with multiple sequentially toggled light sources, the physics of the intensity ratio can be exploited using creative hardware configurations.

Flash / No-Flash Photography

You can achieve albedo invariance using a standard camera simply by taking two photos: one with the camera's built-in flash and one without. Xu Cao et al. (Stereoscopic Flash and No-Flash Photography for Shape and Albedo Recovery) demonstrated that computing the ratio between a no-flash image and a flash-only image mathematically cancels out the surface albedo. This isolates the high-frequency shading gradients caused by the point-source flash, preventing intricate surface patterns from being misinterpreted as physical depth features.

Event Cameras and Logarithmic Invariance

Event cameras do not capture full image frames; instead, they record asynchronous "events" whenever a pixel detects a change in log-intensity ($\Delta \log I$). This unconventional sensing mechanism turns out to be a perfect hardware match for albedo invariance.

As Bohan Yu et al. explain in EventPS: Real-Time Photometric Stereo Using an Event Camera, subtraction in logarithmic space is mathematically identical to division in linear space:

$$\Delta \log I = \log(I_1) - \log(I_2) = \log\left(\frac{I_1}{I_2}\right)$$

Because the static diffuse albedo $\rho$ is constant over time, it disappears during this automatic differentiation. By leveraging a moving light source and the high temporal resolution of event data, the authors can perform real-time, high-resolution photometric stereo that is inherently immune to complex surface textures (EventPSR: Surface Normal and Reflectance Estimation from Photometric Stereo Using an Event Camera).

Alternative Invariants: Statistics and Alignment

When multiple calibrated light sources are unavailable, researchers look for alternative mathematical invariants to decouple shape from reflectance.

  • Statistical Ratios: Daniel Cabrini Hauagge et al. (Photometric Ambient Occlusion) proved that if you observe a surface under a large number of unknown, random lighting conditions, you can calculate a per-pixel statistical ratio: the mean intensity squared divided by the mean of the squared intensities ($\langle I \rangle^2 / \langle I^2 \rangle$). This ratio isolates local geometric visibility (ambient occlusion) while remaining completely invariant to the underlying surface albedo.
  • Angular Alignment: Instead of looking at absolute intensity values, the authors of Photometric refinement of depth maps for multi-albedo objects formulated surface normal estimation by optimizing for the angle between surface vectors and light intensities rather than their magnitudes. Because albedo acts as a purely scalar multiplier, it changes the magnitude of the observed light but leaves the underlying angular alignment intact, enabling robust shape refinement on highly intricate, multi-colored objects.
  • Light-Field Invariants: For non-Lambertian, glossy materials, simple intensity ratios break down because the reflectance changes depending on the viewing angle. To bypass this, Ting-Chun Wang et al. (SVBRDF-Invariant Shape and Reflectance Estimation from Light-Field Cameras) derived a higher-dimensional mathematical invariant within light-field data. By analyzing how light rays shift across sub-aperture views, they successfully decoupled shape estimation from the object's specific, spatially varying specular reflectance profile.

By shifting the focus from explicit material modeling to physics-based invariants, these diverse approaches demonstrate that the most robust way to handle the complexity of real-world textures is not to solve for them, but to design equations where they simply cancel out.

Key Foundational Work

Recovering Shading from Color Images (1992) — contributed 1 concept elements to this direction.
Object-centered surface reconstruction: Combining multi-image stereo and shading. (1995) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 45 papers
  2. Version 1 · 2026-10-04 · 62 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_albedo_invariant_shape_recovery_590d6f3c,
  title = {albedo-invariant shape recovery},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/albedo-invariant-shape-recovery/rdr_590d6f3cb4f34649a06264038d9465d4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_albedo_invariant_shape_recovery_590d6f3c_v2,
  title = {albedo-invariant shape recovery},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/albedo-invariant-shape-recovery/rdr_590d6f3cb4f34649a06264038d9465d4/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake