Scalable MAV Indoor Reconstruction with Neural Implicit Surfaces

Many previous works achieved impressive reconstruction results on room-scale indoor scenes from multi-view RGB images, but capturing and reconstructing multistory, complex indoor scenes is still a challenging problem. In this paper, we propose a fully automated pipeline for reconstructing large and complex indoor scenes with dronecaptured RGB images. First, we leverage traditional structure-from-motion methods to obtain camera poses and reconstruct an initial point cloud. Next, we devise a divide-and-conquer strategy to utilize neural surface reconstruction under the Manhattan-world assumption. Our method reduces the point cloud's outliers and significantly improves reconstruction quality on low-textured regions. We simultaneously predict point-wise semantic logits for walls, floors, and ceilings. The semantic segmentation enables category-wise plane fitting and improves reconstruction quality on polygonal geometry. To validate our method, we use a drone to capture videos inside a large-scale, complex indoor scene. Experimental results showed our method achieved better PSNR in view synthesis tasks and higher floor plan IOU than traditional reconstruction solutions such as COLMAP.