# MeshGS: Adaptive Mesh-Aligned Gaussian Splatting for High-Quality Rendering

**Authors:** Jaehoon Choi, Yonghan Lee, Hyungtae Lee, Heesung Kwon, Dinesh Manocha  
**arXiv:** [2410.08941](https://arxiv.org/abs/2410.08941)  
**Date:** 2024-10-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2410.08941v1/

## Key points

- Paper introduces MeshGS: Adaptive Mesh-Aligned Gaussian Splatting for High-Quality Rendering
- Combines mesh representation with 3D Gaussian splats to enhance rendering quality of real-world scenes
- Focuses on aligning Gaussian splats with mesh surface through distance-based technique to remove redundant splats
- Distinguishes between tightly-bound and loosely-bound splats based on proximity to mesh surface for better alignment and integration with geometry
- Proposes training strategy that binds Gaussian splats to mesh geometry, incorporating both types in rendering process
- Introduces regularization techniques to align tightly-bound splats with mesh surface during training for improved rendering quality
- Validated on large and unbounded scenes from mip-NeRF 360 and Deep Blending datasets, achieving higher PSNR values than recent mesh-based neural rendering techniques
- Comparable rendering quality to existing methods such as original 3DGS and SuGaR in quantitative comparisons on Deep Blending datasets
- Visualization results in Figure 3 demonstrate how MeshGS enhances rendering by reducing redundant Gaussian splats and improving alignment with mesh surface
- Presents significant advancement in high-quality rendering by integrating mesh representation with 3D Gaussian splatting, offering superior results compared to traditional approaches

## Summary

The paper titled "MeshGS: Adaptive Mesh-Aligned Gaussian Splatting for High-Quality Rendering" introduces a novel approach that combines mesh representation with 3D Gaussian splats to enhance the rendering quality of reconstructed real-world scenes. The method focuses on aligning Gaussian splats with the mesh surface through a distance-based technique, effectively removing redundant splats that do not contribute to the final rendering. By distinguishing between tightly-bound and loosely-bound splats based on their proximity to the mesh surface, the method ensures better alignment and integration with the geometry. The authors propose a training strategy that binds Gaussian splats to the mesh geometry, incorporating both types of splats in the rendering process. Several regularization techniques are introduced to precisely align tightly-bound splats with the mesh surface during training, improving overall rendering quality. The effectiveness of the method is validated on large and unbounded scenes from mip-NeRF 360 and Deep Blending datasets, surpassing recent mesh-based neural rendering techniques by achieving higher PSNR values. In quantitative comparisons on Deep Blending datasets, the proposed method demonstrates comparable rendering quality to existing methods such as original 3DGS and SuGaR. Visualization results in Figure 3 showcase how MeshGS enhances rendering by addressing artifacts in reconstructed 3D meshes through a reduction in redundant Gaussian splats and improved alignment with the mesh surface. Overall, MeshGS presents a significant advancement in high-quality rendering by integrating mesh representation with 3D Gaussian splatting, offering superior results compared to traditional approaches. The method's ability to reduce unnecessary splats while enhancing alignment contributes to improved rendering quality across various real-world scenes and datasets.