UDGS-SLAM : UniDepth Assisted Gaussian Splatting for Monocular SLAM

2024-08-31

Array

AI Summary

This work addresses the inherent trade-off between geometric accuracy and rendering quality in monocular SLAM. We propose a joint optimization framework that tightly integrates monocular depth estimation with 3D Gaussian Splatting. Our key contributions are: (1) the first incorporation of the lightweight UniDepth network as a geometric depth prior into Gaussian optimization—without end-to-end coupling; (2) depth-guided adaptive Gaussian initialization and gradient masking to align geometry and appearance; (3) depth-consistency regularization to enforce geometric fidelity; and (4) a keyframe-driven dynamic Gaussian insertion and pruning mechanism for efficient scene representation. Evaluated on TUM and ScanNet, our method achieves state-of-the-art pose accuracy (ATE RMSE < 0.012 m) and real-time rendering at 45 FPS—substantially outperforming MonoGS and iMAP in both localization precision and visual quality.

Technology Category

• Search and Optimization: Sampling/Simulation-based Search
• Computer Vision: Learning & Optimization for CV
• Intelligent Robots: State Estimation

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation

Problem

• Integrates UniDepth with Gaussian splatting for monocular SLAM
• Eliminates need for RGB-D sensors in depth estimation
• Optimizes camera trajectory and scene representation jointly

Innovation

• UniDepth integrated with Gaussian splatting
• Statistical filtering for depth consistency
• Joint optimization of trajectory and scene