Differentiable Point-based Inverse Rendering

TL;DR

A hybrid point-volumetric representation for geometry and a regularized basis-BRDF representation for reflectance that enables fast rendering through point-based splatting while retaining the geometric details and stability inherent to SDF-based representations is devised.

Abstract

We present differentiable point-based inverse rendering, DPIR, an analysis-by-synthesis method that processes images captured under diverse illuminations to estimate shape and spatially-varying BRDF. To this end, we adopt point-based rendering, eliminating the need for multiple samplings per ray, typical of volumetric rendering, thus significantly enhancing the speed of inverse rendering. To realize this idea, we devise a hybrid point-volumetric representation for geometry and a regularized basis-BRDF representation for reflectance. The hybrid geometric representation enables fast rendering through point-based splatting while retaining the geometric details and stability inherent to SDF-based representations. The regularized basis-BRDF mitigates the ill-posedness of inverse rendering stemming from limited light-view angular samples. We also propose an efficient shadow detection method using point-based shadow map rendering. Our extensive evaluations demonstrate that DPIR outperforms prior works in terms of reconstruction accuracy, computational efficiency, and memory footprint. Furthermore, our explicit point-based representation and rendering enables intuitive geometry and reflectance editing.

Figures (12)

• Figure 1: Overview of differentiable forward rendering. (a) For each 3D point, its position is used as a query for the diffuse-albedo MLP $\Theta_d$, SDF MLP $\Theta_\text{SDF}$, and specular-basis coefficient MLP $\Theta_c$. The specular-basis BRDF MLP $\Theta_s$ models specular-basis reflectance, given the incident and outgoing directions $\boldsymbol{\omega_{i}}$ and $\boldsymbol{\omega_{o}}$. The point-based shadow renderer estimates the point visibility from a light source per each image. By using the diffuse albedo, normals, specular reflectance, and visibility, we compute the radiance for each point. (b) The radiance is then projected onto a camera plane to render the pixel color through splatting-based differentiable forward rendering.
• Figure 2: Estimated spatially-varying BRDFs. DPIR accurately reconstruct the BRDFs of the specular gold appearance and the red diffuse appearance. We visualize the BRDFs on unit spheres illuminated by a point light source.
• Figure 3: Point-based visibility test. To determine the visibility of each point, we compute the depth using the z-buffer from a virtual camera positioned at the light source, and compare the depth with the distance from each point to the virtual camera.
• Figure 4: Reconstruction results. DPIR enables accurate and efficient reconstruction of geometry and reflectance for various objects.
• Figure 5: Comparison of novel view rendering and estimated normal on DiLiGenT-MV dataset. Our DPIR recovers detailed surface normals and reproduces accurate appearance.