PlanarSplatting: Accurate Planar Surface Reconstruction in 3 Minutes

TL;DR

PlanarSplatting tackles fast, accurate indoor scene reconstruction by optimizing a set of explicit 3D planar primitives directly from multi-view images. It introduces a rectangle-based, differentiable plane rendering pipeline that splats 3D planes into 2.5D depth and normal maps, guided by monocular depth/normal cues and without relying on plane detection or matching. The method incorporates plane splitting and merging, an effective double radii representation, and a CUDA-accelerated implementation to achieve reconstruction in about 3 minutes per scene, with strong geometry metrics on ScanNet datasets. Moreover, PlanarSplatting can be integrated with Gaussian Splatting methods to improve novel view synthesis efficiency and quality, highlighting the practical impact of a fast, planar-centric representation for indoor scene understanding.

Abstract

This paper presents PlanarSplatting, an ultra-fast and accurate surface reconstruction approach for multiview indoor images. We take the 3D planes as the main objective due to their compactness and structural expressiveness in indoor scenes, and develop an explicit optimization framework that learns to fit the expected surface of indoor scenes by splatting the 3D planes into 2.5D depth and normal maps. As our PlanarSplatting operates directly on the 3D plane primitives, it eliminates the dependencies on 2D/3D plane detection and plane matching and tracking for planar surface reconstruction. Furthermore, the essential merits of plane-based representation plus CUDA-based implementation of planar splatting functions, PlanarSplatting reconstructs an indoor scene in 3 minutes while having significantly better geometric accuracy. Thanks to our ultra-fast reconstruction speed, the largest quantitative evaluation on the ScanNet and ScanNet++ datasets over hundreds of scenes clearly demonstrated the advantages of our method. We believe that our accurate and ultrafast planar surface reconstruction method will be applied in the structured data curation for surface reconstruction in the future. The code of our CUDA implementation will be publicly available. Project page: https://icetttb.github.io/PlanarSplatting/

Figures (11)

• Figure 1: Representation of our 3D plane primitive with learnable shape parameters including plane center, plane radii, and plane rotation.
• Figure 2: Illustration of our proposed PlanarSplatting. Given a set of posed multi-view images of indoor scenes, our method renders depth and normal maps from 3D plane primitives. Then, with the supervision of monocular cues, these 3D plane primitives are gradually optimized to recover the scene geometry and finally merged to get the planar reconstruction result.
• Figure 3: Illustration of the proposed plane splatting function. Naive Gaussian Splatting can not effectively approximate the boundary of our rectangular plane primitive (shown in black dashed border). In contrast, our proposed plane splatting function can approximate the boundary of the rectangle as the number of iterations increases, allowing our 3D planar primitives to better fit the surface of the scene.
• Figure 4: Reconstruction comparison with different splatting functions. 'w/' means 'with'.
• Figure 5: Qualitative comparison on the ScanNetV2 (rows 1-4) and ScanNet++ (last row) datasets.