Multi-view Urban Scene Reconstruction with Continuous Adjustment

Accurate 3D urban scene reconstruction from a set of multi-view images is an important and challenging task in various fields such as computer visions and computer graphics. The previous methods are usually in a dilemma when considering the problems of completeness and accuracy, especially for the complex outdoor scenes. To address these issues, in this paper we propose a new framework based on a number of available methods. Apart from joint optimization with image semantic segmentation, we also attempt to introduce a unified evaluation approach for 3D reconstruction. We will fill holes and remove noise with the guidance of semantic information and the evaluation. The aim of our framework is to improve the quality of urban scene reconstruction in terms of completeness and accuracy for point clouds, which are obtained with state-of-art methods. We demonstrate the effectiveness of our framework on several real public available datasets. As shown in the experimental results, our proposed framework can obtain a better reconstructed model than state-of-art methods.