Skip to main content
Usami Laboratory
Research explainers

Separating Shape and Reflectance from a Single Image

A self-supervised method estimating scene properties like reflectance and shape from a single photo using relighting constraints.

Proceedings of the 60th Tokai Fuzzy Workshop in Toyohashi (Gamaken 2026) · 2026

  1. 01

    Training: lighting conditions

  2. 02

    Learn physical consistency

  3. 03

    Inference: one image

Conceptual overview of the research workflow

The problem

Image brightness mixes material, surface orientation, illumination and distance. Reproducing appearance alone does not establish that geometry and reflectance were separated correctly.

The research idea

Training uses light position, intensity and relit images in a physical reconstruction constraint. At inference, one image is sufficient and privileged lighting information is not used.

What was evaluated

Synthetic images with known scene properties were evaluated on objects absent from training. Privileged-lighting training reduced reflectance, normal and depth errors, while image reconstruction error increased slightly.

Considering applications

Considered for technical exploration in single-image scene understanding and virtual relighting.

Original research

Reflectance Decomposition and 3D Shape Reconstruction via Self-Supervised Single-Image Inverse Rendering with Relighting Constraints

Hara, Keisuke, Usami, Hiroyasu

Proceedings of the 60th Tokai Fuzzy Workshop in Toyohashi (Gamaken 2026), 2026 · S3-03, pp. 15-16, Best Presentation Award, August 19, 2026

Publication recordPublic source