Most computer vision algorithms for 3D reconstruction and material estimation rely on a fundamental simplification: the Bidirectional Reflectance Distribution Function (BRDF). The BRDF assumes that light interacting with an object strikes a surface element and immediately reflects from that exact same point. This local assumption underpins classic shape-from-shading, photometric stereo, and early neural radiance fields.
However, as noted by Goel et al. and Boss et al., this assumption completely breaks down for translucent materials like skin, wax, jade, or marble. When light hits a non-opaque material, it penetrates the surface boundary, undergoes multiple scattering events within the internal medium, and exits from a completely different location. This phenomenon is known as subsurface scattering (SSS).
Treating a translucent object as an opaque surface introduces severe geometric and material errors. Because light bleeds across a neighborhood of pixels before exiting, subsurface scattering acts as a physical spatial low-pass filter. This blurring effect obscures high-frequency visual signals—such as sharp shadows and fine textures—that algorithms use to infer geometry. As Dong et al. and Guo demonstrate, standard photometric stereo and depth scanners fail to distinguish between surface reflection and internal light diffusion, causing them to underestimate surface curvature and produce artificially flattened or distorted 3D shapes.
To overcome this challenge, researchers have developed techniques to untangle surface geometry from internal light transport by isolating uncorrupted signals, mathematically deblurring the scattering effects, or leveraging neural approximations of volumetric transport.
Exploiting the Surface Boundary via Specularity
If subsurface scattering corrupts the diffuse light signal, one powerful intuition is to separate the light that never enters the material from the light that does.
When light strikes a translucent object, a fraction of it is immediately reflected at the air-material interface due to the index of refraction mismatch. This surface reflection produces specular highlights. Crucially, because this light never penetrates the surface, it does not undergo subsurface scattering and remains completely local.
While traditional 3D scanning methods often treat shiny specular highlights as noise to be discarded, Chen et al. inverted this paradigm. By explicitly tracking how specular glints move across a surface, the authors successfully reconstructed the fine-scale "mesostructure" (such as pores and micro-wrinkles) of highly translucent materials like candy and human skin. The specularity acts as an uncorrupted geometric anchor, bypassing the subsurface blur entirely.
Mathematical Deblurring and Multi-Frequency Illumination
Another strategy is to accept that subsurface scattering occurs, but model it explicitly as a convolution operation that can be mathematically inverted. If the internal light transport behaves like a spatial blur, the image can theoretically be "deblurred" to reveal the underlying structure.
Dong et al. formulated this problem as a blind deconvolution task within a photometric stereo framework. By simultaneously optimizing for both the material's scattering parameters and the surface normals, they effectively reversed the low-pass filtering effect of the material to recover sharp surface details.
To resolve ambiguities at different depths, Tanaka et al. introduced multi-frequency illumination. By projecting high-frequency checkerboard patterns of varying sizes onto translucent objects, they analyzed how the material attenuates different spatial frequencies. High-frequency patterns are blurred out rapidly by shallow scattering, whereas lower-frequency patterns penetrate deeper. By measuring this frequency-dependent decay, the authors mathematically "un-mixed" the light transport, allowing them to reconstruct sharp images of distinct internal slices buried beneath the surface.
Differentiable Scattering and Neural Fields
With the rise of differentiable rendering, researchers have moved toward optimizing full 3D representations by minimizing the error between rendered images and real photos. However, simulating subsurface scattering via brute-force Monte Carlo path tracing requires tracking millions of recursive light bounces inside a volume, which is computationally prohibitive for inverse rendering.
To bridge this gap, modern approaches combine physics-based constraints with neural approximations. For instance, Yu et al. introduced Object-Centric Neural Scattering Functions (OSFs). Instead of tracing individual light paths, OSFs train a neural network to model the "cumulative radiance transfer" from any entry point to any exit point on an object, drastically accelerating free-viewpoint relighting.
Similarly, the field has recently adapted explicit point-based representations to handle non-opaque media. Dihlmann integrated subsurface scattering into 3D Gaussian Splatting by pairing explicit physically-based material maps with a lightweight neural network for internal light transport, achieving real-time editing of translucent objects. Zheng et al. expanded on this by introducing SSD-GS, a framework that explicitly decomposes light-material interactions into diffuse, specular, shadow, and subsurface scattering components, allowing photorealistic relighting of complex materials like wax and jade.
Discovering Translucency Priors
When lighting conditions are completely unknown, recovering scattering parameters becomes highly underdetermined. To solve this, researchers look for statistical regularities in how translucent materials behave under varying illumination.
Guo discovered an "inter-pixel translucency prior" based on the observation that when a translucent material is exposed to changing light, the color vectors of neighboring pixels trace out highly distinct, predictable curves in RGB space. By training a neural network (IPTNet) to recognize these geometric trajectories, the system can infer complex internal scattering parameters from just a few casual smartphone photos, turning what was once a complex laboratory calibration task into a practical, mobile-friendly inverse rendering pipeline.
Key Foundational Work
The Kubelka-Munk Theory for Color Image Invariant Properties (2002) — contributed 2 concept elements to this direction.
Mesostructure from Specularity (2006) — contributed 1 concept elements to this direction.