GEVO: Memory-Efficient Monocular Visual Odometry Using Gaussians

Dasong Gao, Peter Zhi Xuan Li, Vivienne Sze, Sertac Karaman

Abstract

Recent advances in 3D Gaussian Splatting (GS) have enabled high-fidelity dense SLAM, but at the cost of significant memory overhead due to the need to store raw keyframes to prevent "catastrophic forgetting." We present GEVO, a memory-efficient monocular Visual Odometry approach that reduces memory usage by up to 94× (down to ~58 MB). GEVO achieves this by rendering past views directly from the map for optimization and employing novel initialization and selective update strategies.

Introduction

Dense Simultaneous Localization and Mapping (SLAM) is critical for AR/VR and robotics. While 3D Gaussian Splatting provides photorealistic reconstructions, existing monocular GS-SLAM frameworks require gigabytes of storage for image buffers. This memory bottleneck limits deployment on energy-constrained edge devices. GEVO addresses this by treating the map itself as the reference for past views, breaking the linear dependency between exploration time and memory consumption.

Methodology

GEVO introduces two primary mechanisms to maintain map integrity without raw image storage:

Occupancy-Preserving Initialization

Standard systems often initialize Gaussians only on visible surfaces, leading to "ghost" artifacts or incorrect occlusions when the camera moves. GEVO utilizes a modified SPGF* algorithm to initialize Gaussians in both occupied and free space. This explicit modeling allows the system to identify and prune Gaussians that cause retrospective occlusions.

Illustration of retrospective occlusion where new Gaussians incorrectly block views of previously mapped surfaces.

Consistency-Aware Optimization

To prevent the map from degrading as new data arrives, GEVO employs a two-stage optimization process. The Local Stage ensures high accuracy for the current sliding window. The Global Stage then merges this data using Selective Optimization. Instead of updating the entire map, the system identifies an "Active Set" of Gaussians that are visible and inconsistent with the current state, preventing the model from overfitting to a single viewpoint.

Conceptual diagram of overfitting, where Gaussians are optimized to match a specific viewpoint at the expense of global consistency.

Experimental Results

Evaluated on the Replica and TUM RGB-D datasets, GEVO demonstrates state-of-the-art memory efficiency:

  • Massive Memory Savings: Overhead is reduced to ~58 MB, a 94× improvement over baselines like MonoGS (which can exceed 3 GB).
  • Reconstruction Quality: GEVO maintains competitive PSNR and SSIM scores, proving that rendered views can effectively substitute for raw keyframes.
  • Real-Time Performance: The system maintains high throughput with minimal latency increases (8-10%) compared to less memory-efficient baselines.

Conclusion

GEVO proves that high-fidelity Gaussian-based SLAM is possible on memory-constrained hardware. By leveraging the map as its own supervisor and explicitly modeling free space, GEVO overcomes the traditional trade-off between reconstruction quality and storage requirements. Future work will focus on integrating large-scale loop closure and optimizing GPU rasterization for further speed gains.

Summary updated

BibTeX
@article{gao2025gevo,
  author = {Dasong Gao and Peter Zhi Xuan Li and Vivienne Sze and Sertac Karaman},
  title = {{GEVO: Memory-Efficient Monocular Visual Odometry Using Gaussians}},
  journal = {IEEE Robotics Autom. Lett.},
  year = {2025},
  doi = {10.1109/lra.2025.3534683},
}
Figure