Core Concepts
DeferredGS proposes a decoupled Gaussian splatting representation that enables efficient 3D scene reconstruction, texture and lighting editing, and realistic relighting by applying deferred shading.
Abstract
The paper introduces DeferredGS, a method for decoupling and editing the Gaussian splatting representation using deferred shading. Gaussian splatting models a 3D scene as a set of 3D Gaussians with attributes like position, rotation, scaling, opacity, and spherical harmonic coefficients. However, the original Gaussian splatting representation entangles both texture and lighting information, making separate texture and lighting editing impossible.
To address this, DeferredGS defines additional attributes such as texture parameters (diffuse albedo, roughness, specular albedo) and normal direction on Gaussians. It also models the illumination with a learnable environment map. Importantly, DeferredGS applies deferred shading, which computes shading at the pixel level, resulting in more realistic relighting effects compared to previous methods that use forward shading.
The key components of DeferredGS are:
Normal Field Distillation: DeferredGS jointly trains a NeRF-like network and a Gaussian splatting representation to distill the normal field from the signed distance function onto the Gaussians' learnable normal attributes, enabling more accurate geometry reconstruction.
Deferred Shading: Instead of performing forward shading for each Gaussian, DeferredGS rasterizes geometry and texture attributes into buffer maps and computes shading at the pixel level under the illumination of a learnable environment map. This avoids blending artifacts when rendering Gaussians under novel lighting conditions.
Editing Capabilities: With the decoupled representation, DeferredGS supports geometry editing by deforming the Gaussians based on a deformed mesh proxy, and texture editing by optimizing the Gaussians' texture attributes to fit both the input edited image and randomly rendered images from different viewpoints.
Experiments demonstrate that DeferredGS outperforms previous methods in terms of novel view synthesis, decomposition, and relighting quality.
Stats
The paper does not provide any specific numerical data or statistics in the main text. The quantitative results are presented in the form of evaluation metrics like PSNR, SSIM, LPIPS, and MSE.
Quotes
"To the best of our knowledge, DeferredGS is the first to apply the deferred shading technique to Gaussian splatting, which alleviates blending artifacts of previous methods."
"Experiments show that our DeferredGS produces more faithful decomposition and editing results compared to previous methods."