Geometry-guided 3D Gaussian Splatting

Abstract

This study proposes a novel 3D Gaussian splatting framework that maximizes the stability of scene reconstruction by integrating metric depth priors from UniDepthV2. While recent Gaussian Splatting-based techniques like self-supervised pose-free 3D Gaussian splatting (SPFSplat) have yielded remarkable results, they still suffer from geometric instability during early training and inconsistent performance in complex scenes. To address these issues, we designed a Depth Module that effectively merges precise depth features extracted from the UniDepthV2 model into the transformer-based backbone of SPFSplat. By employing a zero- convolution-based injection structure to preserve the pretrained feature space and by incorporating camera intrinsic tokens, our framework ensures stable geometric guidance. Experimental results on the RE10K dataset demonstrate that our integrated framework exhibits improved convergence stability during the early training phase (20000–60000 steps) compared with the baseline and attains visual fidelity and geometric consistency comparable to the baseline upon final convergence (300000 steps).