Novel View Pose Synthesis with Geometry-Aware Regularization for Enhanced 3D Gaussian Splatting

@POSTECH Computer Graphics Lab

Project Goal

Enhance the quality of 3D reconstruction
Improve multi view consistency
Incorporate geometry-aware loss terms for accurate surface reconstruction

Project Page

Detailed information about the project can be found in the project page above!

Project Overview

⇣

I developed a method to enhance indoor 3D reconstruction with 3D Gaussian Splatting (3DGS) by generating novel view camera poses, refining them with DIFIX, and applying geometry-aware loss terms. This approach improved geometry accuracy, multi-view consistency, and reduced artifacts.

Contributions

Novel view camera pose generation
- Expanded spatial coverage and ensured consistency between viewpoints.
- Removed artifacts in scenes rendered from novel view camera poses using DIFIX.
Introduction of additional loss terms
- Added a perceptual LPIPS loss applied only to novel views to preserve not only pixel information but also structural details.
- Applied normal consistency loss and depth smoothness loss to all views to improve geometry reconstruction quality.

Results

method	initial point#	PSNR↑	SSIM↑	Training time	frame#
3DGS	100000	20.423	0.856	2h 13m	168
2DGS	100000	19.219	0.828	2h 1m	168
2DGS_novel	100000	20.375	0.842	1h 59m	208
Ours_novel	100000	21.605	0.861	2h 6m	208
Ours_novel_loss	100000	21.675	0.862	3h 55m	208

Compared to 3DGS, our method achieved a PSNR improvement from 20.423 to 21.675 and an SSIM increase from 0.856 to 0.862.
Applying our method to 2DGS also yielded higher scores, demonstrating its generalizability.

🧑‍💻 My Role: Conceived the research idea, designed the methodology, and carried out the entire implementation — including dataset preparation, novel view generation, loss function integration, and experimental evaluation — with advisory input from a doctoral researcher.

GitHub 3dgs-quality-enhancement