CG-SLAM: Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field
(2403.16095)Abstract
Recently neural radiance fields (NeRF) have been widely exploited as 3D representations for dense simultaneous localization and mapping (SLAM). Despite their notable successes in surface modeling and novel view synthesis, existing NeRF-based methods are hindered by their computationally intensive and time-consuming volume rendering pipeline. This paper presents an efficient dense RGB-D SLAM system, i.e., CG-SLAM, based on a novel uncertainty-aware 3D Gaussian field with high consistency and geometric stability. Through an in-depth analysis of Gaussian Splatting, we propose several techniques to construct a consistent and stable 3D Gaussian field suitable for tracking and mapping. Additionally, a novel depth uncertainty model is proposed to ensure the selection of valuable Gaussian primitives during optimization, thereby improving tracking efficiency and accuracy. Experiments on various datasets demonstrate that CG-SLAM achieves superior tracking and mapping performance with a notable tracking speed of up to 15 Hz. We will make our source code publicly available. Project page: https://zju3dv.github.io/cg-slam.
Overview
-
CG-SLAM introduces an efficient dense RGB-D SLAM system using a novel 3D Gaussian field for enhanced speed and accuracy.
-
The system uses Gaussian Splatting for surface modeling and includes innovations for stable tracking and mapping.
-
It outperforms existing NeRF-based SLAM and Gaussian-based SLAM systems in accuracy, reconstruction quality, and efficiency.
-
CG-SLAM's advancements offer significant improvements for applications in virtual/augmented reality, autonomous driving, and robotic navigation.
CG-SLAM: Enhancing Dense RGB-D SLAM with a 3D Gaussian Field
Introduction
Dense RGB-D SLAM technology has achieved notable successes in simultaneous localization and mapping (SLAM) by leveraging neural radiance fields (NeRF) for 3D representations. However, NeRF-based methods suffer from computational inefficiency, hindering their broader application in real-time systems. To address this, CG-SLAM introduces an efficient SLAM system based on a novel consistent and uncertainty-aware 3D Gaussian field. This approach leverages Gaussian Splatting techniques for surface modeling, achieving superior tracking and mapping performance with enhanced speed up to 15 Hz, a significant improvement over conventional methods.
Approach
CG-SLAM employs several innovative techniques to construct a stable and consistent 3D Gaussian field, suitable for accurate tracking and efficient mapping:
- Derivatives of Camera Poses in 3D Gaussian Splatting: A comprehensive mathematical analysis ensures that tracking and mapping processes are decoupled effectively, boosting the overall system's performance.
- Consistent and Stable 3D Gaussian Field: Through scale regularization and alignment of median depth with alpha-blending depth, the system ensures Gaussian primitives are well distributed over scene surfaces, improving geometric stability.
- Depth Uncertainty Model: A novel model improves system accuracy by focusing on stable and valuable Gaussian primitives during optimization, effectively filtering out outliers.
Experimental Evaluation
CG-SLAM was extensively tested across various datasets, including Replica, TUM, and ScanNet, demonstrating superior performance in tracking accuracy, reconstruction quality, and runtime efficiency. Compared to state-of-the-art NeRF-SLAM methods and concurrent Gaussian-based SLAM systems, CG-SLAM exhibits competitive or better results, highlighting its effectiveness in practical applications.
Implications and Future Work
The introduction of CG-SLAM opens new possibilities for efficient and accurate dense SLAM applications. Its superior performance could significantly benefit virtual/augmented reality, autonomous driving, and robotic navigation. Future developments could explore more compact representations of the 3D Gaussian field to reduce memory usage and improve system's predictability in unobserved areas. Additionally, extending the system to handle dynamic environments presents an interesting direction for further research.
Conclusion
CG-SLAM represents a significant advancement in dense RGB-D SLAM by integrating a consistent and efficient 3D Gaussian field. Its capability to achieve high tracking accuracy, detailed reconstruction, and real-time performance positions it as a powerful tool for various 3D computer vision applications. With room for further improvements and adaptations, CG-SLAM sets a new benchmark for future research in the field of SLAM technology.