RayGauss: Volumetric Gaussian-Based Ray Casting for Photorealistic Novel View Synthesis

TL;DR

RayGauss tackles the challenge of photorealistic novel view synthesis by representing radiance fields with irregular, anisotropic Gaussian basis functions and a direction-dependent radiance model. It introduces a slab-by-slab volume ray casting scheme accelerated by a GPU BVH (OptiX) to enable differentiable rendering with efficient training and real-time-like inference. The approach achieves state-of-the-art quality on Blender and Mip-NeRF360 datasets, with average PSNR improvements and 25 FPS inference on Blender, while using a compact yet expressive parameterization (SH/SG) to capture both low- and high-frequency appearance. Ablation studies show that anisotropic Gaussians and the combination of SH and SG yield the strongest performance, though training speed remains a consideration for highly irregular primitive sets; future work could extend to light scattering and surface-level relighting.

Abstract

Differentiable volumetric rendering-based methods made significant progress in novel view synthesis. On one hand, innovative methods have replaced the Neural Radiance Fields (NeRF) network with locally parameterized structures, enabling high-quality renderings in a reasonable time. On the other hand, approaches have used differentiable splatting instead of NeRF's ray casting to optimize radiance fields rapidly using Gaussian kernels, allowing for fine adaptation to the scene. However, differentiable ray casting of irregularly spaced kernels has been scarcely explored, while splatting, despite enabling fast rendering times, is susceptible to clearly visible artifacts. Our work closes this gap by providing a physically consistent formulation of the emitted radiance c and density σ, decomposed with Gaussian functions associated with Spherical Gaussians/Harmonics for all-frequency colorimetric representation. We also introduce a method enabling differentiable ray casting of irregularly distributed Gaussians using an algorithm that integrates radiance fields slab by slab and leverages a BVH structure. This allows our approach to finely adapt to the scene while avoiding splatting artifacts. As a result, we achieve superior rendering quality compared to the state-of-the-art while maintaining reasonable training times and achieving inference speeds of 25 FPS on the Blender dataset. Project page with videos and code: https://raygauss.github.io/

Figures (4)

• Figure 1: Our method (RayGauss) compared to the main radiance field rendering methods
• Figure 2: Test set images from the Mip-NeRF 360 Dataset
• Figure 3: Overview of our slab-by-slab Ray Casting algorithm: When a given ray intersects the scene's bounding box, the ray is processed by successive slabs. For a given slab, we check the intersection with the Gaussians using the BVH. Then, the attributes of the slab are calculated to update the current color and transmittance of the ray.
• Figure 4: Test set images from the Blender Dataset