OG-Mapping: Octree-based Structured 3D Gaussians for Online Dense Mapping

TL;DR

OG-Mapping tackles the challenge of robust, real-time online dense mapping by fusing a sparse octree with structured 3D Gaussians anchored to scene voxels. It introduces an anchor-based progressive map refinement and a dynamic keyframe window to improve detail capture, avoid local minima, and mitigate forgetting, all while maintaining a compact memory footprint. Across Replica and ScanNet, it delivers higher rendering fidelity and more compact maps than prior Gaussian-based RGB-D methods, benefiting from fast rasterization and depth-noise robustness. The approach advances real-time 3D mapping for AR/VR and digital twins by enabling high-quality dense reconstructions with efficient memory usage.

Abstract

3D Gaussian splatting (3DGS) has recently demonstrated promising advancements in RGB-D online dense mapping. Nevertheless, existing methods excessively rely on per-pixel depth cues to perform map densification, which leads to significant redundancy and increased sensitivity to depth noise. Additionally, explicitly storing 3D Gaussian parameters of room-scale scene poses a significant storage challenge. In this paper, we introduce OG-Mapping, which leverages the robust scene structural representation capability of sparse octrees, combined with structured 3D Gaussian representations, to achieve efficient and robust online dense mapping. Moreover, OG-Mapping employs an anchor-based progressive map refinement strategy to recover the scene structures at multiple levels of detail. Instead of maintaining a small number of active keyframes with a fixed keyframe window as previous approaches do, a dynamic keyframe window is employed to allow OG-Mapping to better tackle false local minima and forgetting issues. Experimental results demonstrate that OG-Mapping delivers more robust and superior realism mapping results than existing Gaussian-based RGB-D online mapping methods with a compact model, and no additional post-processing is required.

Figures (6)

• Figure 1: In this work, we introduce OG-Mapping, a novel online dense mapping framework with an octree-based structured 3D Gaussian representation. By combining our proposed anchor-based progressive map refinement strategy and dynamic keyframe window, OG-Mapping achieves fast, high-fidelity online reconstruction with efficient memory usage, and demonstrates superior realism in novel view synthesis compared to other existing RGB-D online mapping methods. The rendering FPS is indicated to the right of each method.
• Figure 2: Overview of OG-Mapping. Given a set of sequential RGB-D frames and camera poses, we utilize an octree-based structured 3D Gaussians as the scene representation to perform efficient online dense mapping. When a new keyframe is detected, we employ a sparse octree to swiftly capture the rough structure of the new observed region to guide anchor densification (Sec. \ref{['sec: represention']}) . During the map update process, we perform anchor-based progressive map refinement to enhance the geometry and appearance quality (Sec. \ref{['sec: optimization']}), and construct a dynamic keyframe window to effectively mitigate false local minima and forgetting issues (Sec. \ref{['sec: keyframe']}).
• Figure 3: Anchor-based Progressive Map Refinement. We grow new anchors in under-optimized regions based on the level of 3D Gaussians and their gradients. If the regions already contain anchors at the same level, the granularity of these new anchors will be increased.
• Figure 4: Illustration of the differences between Fixed Keyframe Window and our Dynamic Keyframe Window. Left: Fixed Keyframe Window maintains a static window content across all optimization iterations. Right: our dynamic Keyframe Window updates the keyframe window with new data each optimization iteration.
• Figure 5: Qualitative comparison of rendering results across three scenes from the Replica dataset. Key details are highlighted by colored boxes. The average PSNR metric for each scene is indicated in the lower-left corner.