Title: Learning Subsurface Scattering: Acquisition and Representation via Neural Networks

Abstract: This paper introduces a novel approach for capturing and estimating the intricate subsurface scattering characteristics of light transport. By learning the pixel footprint response at individual surface points, the method achieves a high level of detail. The reconstruction process utilizes 3D scanning data as input for a U-Net Convolutional Neural Network (CNN). Data collection is facilitated by a stereo projector-camera system employing phase-shifted profilometry (PSP) patterns, which effectively records information across a range of scattering materials. This technique enables the reconstruction of dense pixel footprints, allowing for the relighting of objects using arbitrary, high-resolution projector patterns to generate a final relit color image. Both qualitative and quantitative evaluations against real-world captured images confirm that the predicted footprints closely match actual responses. Furthermore, the model is trained across multiple views and objects, enabling the learned representations to generalize effectively to previously unseen subsurface scattering materials.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC