Technology Category
Computer Vision: 3D Computer VisionKnowledge Representation and Reasoning: Geometric, Spatial, and Temporal ReasoningIntelligent Robots: Multimodal Perception & Sensor Fusion
CDGS: Confidence-Aware Depth Regularization for 3D Gaussian Splatting
🤖 AI Summary
3D Gaussian Splatting (3DGS) achieves efficient and high-fidelity novel view synthesis but suffers from limited 3D reconstruction accuracy due to the absence of explicit geometric constraints. To address this, we propose a confidence-aware adaptive depth regularization method that integrates monocular depth estimation confidence maps with sparse SfM depth as geometric priors, and introduces a differentiable dynamic-weighted depth loss to enforce geometric fidelity early in training. This mechanism requires no additional networks or ground-truth annotations. Experimental results on Tanks and Temples demonstrate significant improvements: PSNR increases by 2.31 dB, M3C2 geometric error is substantially reduced, and the F-score matches that of standard 3DGS using only 50% of the training iterations—while also achieving more stable convergence.
Technology Category
Application Category
📝 Abstract
Abstract. 3D Gaussian Splatting (3DGS) has shown significant advantages in novel view synthesis (NVS), particularly in achieving high rendering speeds and high-quality results. However, its geometric accuracy in 3D reconstruction remains limited due to the lack of explicit geometric constraints during optimization. This paper introduces CDGS, a confidence-aware depth regularization approach developed to enhance 3DGS. We leverage multi-cue confidence map of monocular depth estimation and sparse Structure-from-Motion (SfM) depth to adaptively adjusts depth supervision during the optimization process. Our method demonstrates improved geometric detail preservation in early training stages and achieves competitive performance in both NVS quality and geometric accuracy. Experiments on the public available Tanks and Temples benchmark dataset show that our method achieves more stable convergence behavior and more accurate geometric reconstruction results, with improvements of up to 2.31 dB in PSNR for NVS and consistently lower geometric errors in M3C2 distance metrics. Notably, our method reaches comparable F-scores to the original 3DGS with only 50% of the training iterations. We expect this work will facilitate the development of efficient and accurate 3D reconstruction systems for real-world applications such as digital twin creation, heritage preservation, or forestry applications.
Problem
Research questions and friction points this paper is trying to address.
Enhance 3D Gaussian Splatting geometric accuracy
Improve novel view synthesis quality and stability
Facilitate efficient 3D reconstruction for real-world applications
Innovation
Methods, ideas, or system contributions that make the work stand out.
Confidence-aware depth regularization
Multi-cue confidence maps integration
Enhanced geometric accuracy in 3DGS
🔎 Similar Papers
No similar papers found.
