Diskchungs: Scalable 3D Gaussian SLAM Enables Large-Scale Reconstruction through Chunk-Based Memory Management

The challenge of creating detailed and expansive 3D maps has long been limited by computer memory, particularly when using the increasingly popular technique of 3D Gaussian Splatting. Now, Casimir Feldmann, Maximum Wilder-Smith, and Vaishakh Patil, from the Robotic Systems Lab at ETH Zurich, alongside colleagues at Google including Michael Oechsle, Michael Niemeyer, and Keisuke Tateno, present a solution with DiskChunGS. This innovative system overcomes memory limitations by intelligently partitioning scenes into manageable chunks, storing inactive areas on disk and only loading what is needed into GPU memory. The result is a scalable 3D Gaussian Splatting SLAM system capable of reconstructing large-scale environments, demonstrated by its successful completion of all eleven KITTI driving sequences, a feat previous methods could not achieve, and its robust performance on indoor and resource-constrained platforms.

It achieves scalability by employing out-of-core processing, leveraging disk storage to handle large-scale reconstructions that would otherwise exceed GPU memory capacity. Key contributions include spatial chunking, where the 3D scene is divided into manageable pieces, and disk-based management, which dynamically loads and unloads chunks as needed to minimize GPU memory usage. The system also incorporates a locality-focused keyframe strategy, optimizing data input/output by prioritizing chunks near current keyframes.

This allows DiskChunGS to reconstruct multi-kilometer environments on standard GPUs, a capability previously unattainable. The system achieves competitive visual quality compared to in-memory 3DGS SLAM systems and demonstrates real-world applicability through integration with ROS and validation on resource-constrained platforms like the Jetson Orin. DiskChunGS treats large-scale 3D reconstruction as an algorithmic challenge rather than a hardware one. It divides the 3D space into manageable chunks, stores them on disk, and dynamically loads and unloads them based on proximity to the current viewpoint and keyframes.

This allows the system to process scenes much larger than what can fit into GPU memory. The paper demonstrates the effectiveness of DiskChunGS through successful reconstruction of scenes spanning multiple kilometers, achieving comparable or better visual quality than existing 3DGS SLAM methods, and delivering real-time performance on standard hardware. Researchers have now presented DiskChunGS, a scalable 3DGS SLAM system that overcomes this bottleneck through an innovative out-of-core approach. This system partitions scenes into spatial chunks, maintaining only active regions in GPU memory while storing inactive areas on disk, effectively expanding the size of reconstructable environments. The core of DiskChunGS lies in its ability to dynamically manage scene data, loading and unloading spatial chunks based on camera visibility, and seamlessly integrating with existing SLAM frameworks for pose estimation and loop closure.

Extensive validation on indoor scenes like Replica and TUM-RGBD, as well as urban driving scenarios using the KITTI dataset, demonstrates the system’s robust performance. Notably, DiskChunGS uniquely completed all 11 KITTI sequences without encountering memory failures, a significant achievement over previous methods. The breakthrough delivers superior visual quality while addressing a critical limitation in large-scale 3D reconstruction. Measurements confirm that the system achieves state-of-the-art performance across diverse datasets, and the architecture scales effectively across hardware platforms, including resource-constrained Nvidia Jetson devices.

Evaluations demonstrate that the system maintains high frame rates and low perceptual image distortion, achieving superior results compared to competing methods like CaRtGS and GigaSLAM. Specifically, the system achieves higher reconstruction quality in less time, as demonstrated by Pareto curves generated during testing. This work introduces a novel out-of-core chunk-based architecture that enables large-scale 3DGS SLAM by partitioning scenes and dynamically managing data between disk and VRAM. Researchers developed an out-of-core system that partitions scenes into spatial chunks, storing inactive areas on disk and managing data efficiently to enable reconstruction at a scale previously unattainable. This approach allows for the creation of detailed 3D models of multi-kilometer environments using standard GPU hardware. Evaluations across diverse scenarios, including indoor spaces, urban driving sequences, and resource-constrained platforms, demonstrate the method’s superior visual quality and scalability.

The system successfully completed all sequences within the challenging KITTI dataset without memory failures, a feat that eluded previous approaches. While the method occasionally exhibits sparse floating artifacts in transitions, and distant objects may lack detail, the team acknowledges these limitations and proposes future work to address them. This includes leveraging GPU Direct Storage to reduce data transfer costs and integrating a Level of Detail system for improved rendering of distant objects, alongside evaluation on even larger, ultra-long sequences to further demonstrate scalability.