Papers
Topics
Authors
Recent
Search
2000 character limit reached

LoopSplat: Loop Closure by Registering 3D Gaussian Splats

Published 19 Aug 2024 in cs.CV and cs.RO | (2408.10154v2)

Abstract: Simultaneous Localization and Mapping (SLAM) based on 3D Gaussian Splats (3DGS) has recently shown promise towards more accurate, dense 3D scene maps. However, existing 3DGS-based methods fail to address the global consistency of the scene via loop closure and/or global bundle adjustment. To this end, we propose LoopSplat, which takes RGB-D images as input and performs dense mapping with 3DGS submaps and frame-to-model tracking. LoopSplat triggers loop closure online and computes relative loop edge constraints between submaps directly via 3DGS registration, leading to improvements in efficiency and accuracy over traditional global-to-local point cloud registration. It uses a robust pose graph optimization formulation and rigidly aligns the submaps to achieve global consistency. Evaluation on the synthetic Replica and real-world TUM-RGBD, ScanNet, and ScanNet++ datasets demonstrates competitive or superior tracking, mapping, and rendering compared to existing methods for dense RGB-D SLAM. Code is available at loopsplat.github.io.

Citations (9)

Summary

  • The paper introduces LoopSplat, a novel RGB-D SLAM system leveraging 3D Gaussian splats for effective loop closure.
  • It presents a fast 3DGS registration method that minimizes tracking errors, achieving an average error as low as 0.26 cm.
  • LoopSplat improves global map consistency and reconstruction quality, validated across synthetic and real-world datasets.

LoopSplat: Loop Closure by Registering 3D Gaussian Splats

Introduction

The paper, LoopSplat: Loop Closure by Registering 3D Gaussian Splats, presents a novel approach to Simultaneous Localization and Mapping (SLAM) with RGB-D cameras. The crux of the proposed method lies in leveraging 3D Gaussian Splats (3DGS) for efficient and accurate dense mapping. The primary challenges identified with existing 3DGS-based SLAM methods include the lack of strategies to achieve global consistency through loop closure and global bundle adjustment. The authors propose LoopSplat, a new SLAM system that directly utilizes 3DGS for executing loop closure, thus enhancing global consistency without the redundancies of saving mapped input frames or relying on traditional point cloud registration techniques.

Key Contributions

  1. Introduction of LoopSplat: The paper introduces LoopSplat, a coupled RGB-D SLAM system with Gaussian Splatting at its core. The system includes a loop closure module that directly operates on Gaussian splats. This integration of both 3D geometry and visual scene content facilitates robust loop detection and closure.
  2. 3DGS Registration Technique: The authors propose a novel method for registering two 3DGS representations efficiently. Utilizing the fast rasterization of 3DGS, the proposed registration method proves superior in speed and accuracy compared to traditional techniques.
  3. Enhanced Tracking and Reconstruction: Through robust loop closure mechanisms and improved pose graph optimization, LoopSplat demonstrates marked improvements in the tracking accuracy and robustness of 3DGS-based RGB-D SLAM systems across diverse real-world datasets.

Methodology

Front-End SLAM

LoopSplat employs submaps to handle local frame-to-model tracking and dense mapping. Each submap represents a segment of the scene with multiple 3D Gaussian components. A new submap is triggered based on predefined thresholds for relative displacement and rotation between frames. Tracking is achieved through frame-to-model alignment via minimizing a tracking loss, which combines color and depth information. Submaps are dynamically expanded by adding new 3D Gaussians sampled based on sparsely observed regions and updated iteratively through a rendering loss function.

3DGS Registration

Registration of overlapping 3D Gaussian submaps is vital for loop closure. The authors propose a method that refrains from traditional point cloud techniques. Instead, they register two 3DGS representations by treating submap viewpoints as rigid bodies. Using selected viewpoint pairs exhibiting high visual similarity, their relative poses are optimized to minimize rendering discrepancies. The final transformation aligning submaps is derived through multi-view pose refinement and weighted rotation averaging, making the process efficient and accurate.

Loop Closure

Loop closure in LoopSplat is conducted online. The system detects loops by calculating the visual similarity between submaps through NetVLAD descriptors and validates these loops using geometric overlap ratios. Once confirmed, loop edge constraints are generated via the aforementioned 3DGS registration method, and global pose graph optimization is executed to achieve globally consistent mapping. This approach ensures continuous correction of pose drift and map deformation throughout the mapping process, leading to high-quality reconstructions.

Experimental Evaluation

Datasets

The evaluation is performed on a synthetic dataset (Replica) and three real-world datasets: TUM-RGBD, ScanNet, and ScanNet++. Each dataset presents unique challenges, from high-quality synthetic scenes to challenging real-world scenarios with significant motion and scene complexity.

Results

Tracking Accuracy: The authors demonstrate superior tracking performance across all datasets. On the synthetic Replica dataset, LoopSplat outperforms all baselines with an average tracking error of 0.26 cm, showcasing its precision. For complex real-world datasets like ScanNet and ScanNet++, LoopSplat shows stable and accurate pose estimates despite large camera motions and scene loops.

Reconstruction: Even in challenging real-world environments, LoopSplat achieves superior reconstruction fidelity. Evaluated through metrics like F1-score and depth error, LoopSplat consistently produces accurate and detailed maps. This is further evidenced by qualitative results demonstrating geometry preservation and detail recovery.

Rendering: For training views, LoopSplat surpasses all competing methods in terms of PSNR and LPIPS across datasets. The method excels in rendering with fewer artifacts and higher texture quality, indicative of superior map quality.

Efficiency: Despite its complexities, LoopSplat maintains competitive runtime and memory efficiency. Particularly, the loop edge registration and overall memory footprint are significantly optimized compared to baseline methods.

Conclusion

LoopSplat presents a robust advancement in RGB-D SLAM by seamlessly integrating 3D Gaussian Splats for dense mapping, tracking, and loop closure. By addressing the challenges of global consistency and map maintenance, the proposed method demonstrates remarkable improvements in speed, accuracy, and memory efficiency. Future research directions may consider the integration of advanced mesh extraction techniques and uncertainty modeling for further enhancing the registration and reconstruction processes.

In summary, this paper introduces a compelling method offering significant strides in SLAM technology, grounded in a solid theoretical framework and validated by extensive empirical evidence. These advancements pave the way for more efficient and scalable SLAM systems capable of operating in diverse and complex environments.

Paper to Video (Beta)

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Collections

Sign up for free to add this paper to one or more collections.

Tweets

Sign up for free to view the 3 tweets with 43 likes about this paper.