# 3D Gaussian Splatting with Deferred Reflection

**Authors:** Keyang Ye, Qiming Hou, Kun Zhou  
**arXiv:** [2404.18454](https://arxiv.org/abs/2404.18454)  
**Date:** 2024-04-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2404.18454v1/

## Key points

- Authors: Keyang Ye, Qiming Hou, and Kun Zhou
- Challenges in rendering specular reflection in novel view synthesis using neural and Gaussian-based radiance field methods
- Proposal of deferred shading method utilizing Gaussian splatting for rendering specular reflections
- Difficulty in accurate specular reflections due to environment map reflection model requiring precise surface normals
- Leveraging per-pixel reflection gradients through deferred shading to optimize neighboring Gaussians for normal estimations
- Improved accuracy and stability in fitting high-frequency radiance fields for specular reflections
- Outperformance of existing techniques and concurrent work in synthesizing high-quality specular reflection effects demonstrated through extensive experimentation
- Consistent improvements in peak signal-to-noise ratio (PSNR) for synthetic and real-world scenes while maintaining comparable frame rate to vanilla Gaussian splatting

## Summary

In their paper titled "3D Gaussian Splatting with Deferred Reflection," authors Keyang Ye, Qiming Hou, and Kun Zhou address the challenges of rendering specular reflection in novel view synthesis using neural and Gaussian-based radiance field methods. These techniques have shown great success but accurately fitting high-frequency radiance fields for specular reflection remains a significant challenge. To overcome this issue, the authors propose a deferred shading method that utilizes Gaussian splatting to effectively render specular reflections. The key obstacle in rendering accurate specular reflections lies in the environment map reflection model which requires precise surface normals. However, estimating normals can be difficult due to discontinuous gradients. To tackle this challenge, the authors leverage per-pixel reflection gradients generated through deferred shading to facilitate the optimization process of neighboring Gaussians. This approach allows for nearly correct normal estimations to gradually propagate and spread across all reflective objects, resulting in improved accuracy and stability in fitting high-frequency radiance fields for specular reflections. Through extensive experimentation, the authors demonstrate that their method outperforms existing techniques and concurrent work in synthesizing high-quality specular reflection effects. They showcase consistent improvements in peak signal-to-noise ratio (PSNR) for both synthetic and real-world scenes while maintaining a frame rate comparable to vanilla Gaussian splatting. Overall, their proposed approach offers a promising solution for efficiently rendering specular reflections in novel view synthesis applications.