Image-based object modeling by fitting salient lines and geometric primitives

Abstract

With modern vision techniques and depth sensing devices, it becomes possible for common users to acquire the shape of an object from a set of color or depth images from different views. However, the estimated 3D volume or point clouds, disturbed by noise and errors, cannot directly be applied for graphics usage. This paper presents a two-stage method for reconstructing 3D graphics models from point clouds and photographs. Unlike related work that immediately fitted primitives for the point clouds, we propose finding the primary planes through salient lines in images in advance, and extracting auxiliary planes according to the symmetric properties. Then, a RANSAC method is used to fit primitives for the residual points. Intuitive editing tools are also provided for rapid model refinement. The experiments demonstrate that the proposed automatic stages can generate more accurate results. Besides, the user intervention time is less than that by a well known modeling tool.