TL;DR
DrawSplat is a hybrid RGB-D SLAM framework that mimics a drawing process: it first "sketches" indoor scene structure using explicit point-line geometry and then "fills" in high-fidelity details with implicit neural Gaussians. By leveraging the Atlanta World prior to regularize planar surfaces, it achieves state-of-the-art tracking and reconstruction quality with a remarkably compact model size (<20MB).
What Problem Does This Paper Address?
While 3D Gaussian Splatting (3DGS) has revolutionized real-time neural rendering, adapting it for Simultaneous Localization and Mapping (SLAM) remains challenging. Most Gaussian SLAM systems rely on local photometric and depth constraints, which often fall short in indoor environments. These scenes are frequently plagued by "weak textures" (like white walls), noisy depth sensor data, and long trajectories that lead to pose drift.
The result? Maps that suffer from "surface thickness" artifacts (where a wall looks like a fuzzy cloud), noisy normals, and blurred structural boundaries. Existing structural priors often introduce heavy computational overhead or struggle to maintain consistency between the geometry and the appearance of the scene.
Methods
DrawSplat introduces a hybrid paradigm that splits the workload between a geometric front-end and a neural back-end:
- Explicit Geometry-Guided Tracking: The system extracts both ORB point features and line segments (using EDLines). This joint point-line tracking provides much more stable "locking" of camera poses in corridors and staircases where points alone might fail. Crucially, it uses these lines to estimate a stable gravity direction.
- Atlanta World (AW) Prior: Unlike the restrictive Manhattan World assumption (which assumes all walls are at 90-degree angles), the Atlanta World prior allows for multiple horizontal directions while sharing a single vertical gravity vector. DrawSplat infers structural planes from tracked lines and uses the AW prior to regularize Gaussian centers and normals through differentiable loss terms.
- Structure-Aware Implicit Representation: Instead of storing millions of individual Gaussians, DrawSplat uses a sparse grid of anchors. Lightweight MLPs decode these anchors into local Gaussian primitives. This ensures spatial continuity and drastically reduces the memory footprint.
- Adaptive Structural Sampling: Rather than densifying the map uniformly, the system allocates "Gaussian capacity" to structurally informative regions—like corners, door frames, and edges—based on line responses and depth gradients.
Datasets
The authors evaluated DrawSplat on four major indoor benchmarks:
- Replica: A high-quality synthetic dataset for controlled evaluation.
- ScanNet: A real-world dataset featuring significant sensor noise and motion blur.
- ScanNet++: Used for qualitative comparisons of complex geometry.
- Replica Apartment: A large-scale, multi-room benchmark designed to test long-range consistency and drift.
Key Insights and Ideas
- The "Drawing" Metaphor: The core "aha" moment is treating SLAM like a sketch. By establishing the "geometric skeleton" (lines and planes) first, the neural optimization of appearance becomes a constrained, well-posed problem rather than a chaotic search.
- Line-Induced Planes: Lines are far more informative than sparse points for recovering indoor structure. DrawSplat converts sparse line observations into structural plane targets, which "pull" Gaussians toward the true surface.
- Semantic Modulation: Not every object follows the Atlanta World prior (e.g., a messy pile of clothes). The system uses semantic labels to apply structural regularization only to "architectural" elements like walls, floors, and ceilings, preventing the over-smoothing of furniture and clutter.
- Efficiency through Sparsity: By combining voxel-based anchors with adaptive sampling, the system achieves a "cleaner" mesh than explicit methods like LoopSplat while using significantly less memory.
Key Results
DrawSplat demonstrates a superior balance of quality and efficiency:
- Tracking Accuracy: On the multi-room Replica Apartment dataset, DrawSplat reduced trajectory error by 60% compared to the next best baseline (MG-SLAM).
- Rendering Quality: It improved rendering fidelity (PSNR) by 2.54 dB on the real-world ScanNet dataset.
- Model Compactness: The system maintains a model size of less than 20 MB, which is significantly smaller than many explicit Gaussian SLAM counterparts (often 100MB+).
- Visual Fidelity: As seen in Figure 3, DrawSplat preserves much finer visual details and sharper geometric boundaries compared to MonoGS or LoopSplat.
Limitations and Future Work
The authors acknowledge that DrawSplat is specifically tailored for structured indoor scenes. Its performance may degrade in:
- Highly dynamic environments where the "static world" assumption fails.
- Outdoor scenes or spaces that do not satisfy the Atlanta World assumption (e.g., organic, non-planar architecture).
- Severe depth errors: While robust, extremely unreliable sensor data can still impair plane estimation.
Future work aims to extend the framework to more general layouts and explore monocular SLAM by integrating metric depth prediction models.
Summary updated