2021/03/20 by Sinisa Stekovic, Mahdi Rad, Stekovic, Sinisa +5 · 2 citations
Engineering · Environmental Science · Earth and Planetary Sciences · #3D Shape Modeling and Analysis #Remote Sensing and LiDAR Applications #3D Surveying and Cultural Heritage
paper · pdf · doi:10.48550/arxiv.2103.11161
We propose a novel method for reconstructing floor plans from noisy 3D point\nclouds. Our main contribution is a principled approach that relies on the Monte\nCarlo Tree Search (MCTS) algorithm to maximize a suitable objective function\nefficiently despite the complexity of the problem. Like previous work, we first\nproject the input point cloud to a top view to create a density map and extract\nroom proposals from it. Our method selects and optimizes the polygonal shapes\nof these room proposals jointly to fit the density map and outputs an accurate\nvectorized floor map even for large complex scenes. To do this, we adapted\nMCTS, an algorithm originally designed to learn to play games, to select the\nroom proposals by maximizing an objective function combining the fitness with\nthe density map as predicted by a deep network and regularizing terms on the\nroom shapes. We also introduce a refinement step to MCTS that adjusts the shape\nof the room proposals. For this step, we propose a novel differentiable method\nfor rendering the polygonal shapes of these proposals. We evaluate our method\non the recent and challenging Structured3D and Floor-SP datasets and show a\nsignificant improvement over the state-of-the-art, without imposing any hard\nconstraints nor assumptions on the floor plan configurations.\n