Physics-based computer vision models the behavior of light and solves inverse problems to recover 3D shape, reflectance, material properties, and illumination from images.
Our laboratory studies reflection analysis, gloss estimation, transparent-object recognition, polarization cues, and inverse rendering.
Why It Is Difficult
Real-world lighting is complex. Interreflections, subsurface scattering, anisotropic reflection, and mixed materials often violate simple image formation models.
The problem is often ill-posed, so uncertainty, ambiguity, and physically plausible constraints must be handled explicitly.
Approach
We study hybrid methods that embed physical constraints into deep learning and optimization-based methods using differentiable rendering.
The goal is to connect data-driven representations with image formation models that remain meaningful outside controlled datasets.
Evaluation
We combine quantitative evaluation on synthetic data with qualitative and consistency-based evaluation on real data.
When ground truth is hard to obtain, we use re-rendering consistency and physically motivated diagnostics.
Current Questions
- 1Reflectance estimation in outdoor or ambient-light environments
- 23D reconstruction including transparent and translucent objects
- 3Illumination estimation from a single image