Reconstructing 3D Contour Models of General Scenes from RGB-D Sequences

General 3D reconstruction methods use voxels, surfels, or meshes to represent the 3D model of a given scene. These surface-based methods are vulnerable to the loss of boundary details, which affects the completeness of the reconstructed model. In this paper, we focus on the boundary information of the scene and propose a novel method to reconstruct 3D models by using 3D contours extracted from input image sequences. We design a robust frame-to-model contour matching algorithm to solve the problem of finding many-to-many contour correspondences between different frames, and use contour-enhanced optimization to obtain more accurate camera poses. In order to make the reconstructed model more expressive of structural information, we propose a contour fusion algorithm that considers the connections between 3D contours. Compared with other methods which use straight lines or curve segments to reconstruct the scene model, our method can generate a more complete and regular 3D contour model with topological relationship. Experiments on several public datasets demonstrate the effectiveness of our method for both modeling and pose estimation.