Coarse-to-fine pipeline for 3D wireframe reconstruction from point cloud

Abstract

Point clouds captured by 3D scans are typically sparse, irregular, and noisy, resulting in 3D wireframes reconstructed by existing approaches often containing redundant edges or lacking proper edges. To tackle these issues, this paper proposes a coarse-to-fine pipeline for 3D wireframe reconstruction from point clouds. First, a learning-based module is dedicated to predicting the corner and edge points from the input point cloud, and each pair of corner points is linked together to generate an initial 3D wireframe. Second, a coarse pruning module is utilized to generate a coarse 3D wireframe, which is achieved by pruning observable redundant edges from the initial 3D wireframe based on the asymmetric Chamfer distance. Third, a refined pruning module is used to generate a refined 3D wireframe with correct topological structures, which can help prune redundant edges that are difficult to observe from the coarse 3D wireframe. Finally, a heuristic algorithm is exploited to fine-tune the refined 3D wireframe to ensure the final 3D wireframe preserves the characteristics of both vertical and parallel. The experimental results reveal that the proposed method significantly improves the performance of 3D wireframes reconstruction from point clouds on the large-scale ABC dataset and a challenging furniture dataset.