An indoor stereo map completion dataset and a novel corresponding method

摘要

It is a common issue to remove undesirable objects during reconstruction. The holes left behind are then needed to be given plausible information for the integrity. To complete the missing part of a 3D map generated from stereo inputs, the task can be separated into two operations which are repairing the color image and the depth map. In this paper, we design and release a specialized stereo dataset for testing and benchmarking 3D map completion algorithms. Different from previous datasets, we give the initial photos and damaged images and disparities by manually taking targets out of the picture. An indoor 3D map completion method is proposed in this paper with the dataset. The proposed method combines an image inpainting method and the planarity knowledge of depth maps. Several real-world examples are presented in the experiment to demonstrate the performance of our 3D map completion technique.