基于图片的交互式建筑物建模

As the wide spread of various camera equipment, image becomes easy to be acquired, yet the research on how to quickly reconstruct 3D model of real world from images is still a hot topic and hard topic in both Computer Graphics and Computer Vision field. The 3D model reconstruction of architectures, as the most important part of the urban scene, attracts wide attention from both academia and industry. Among existing image-based architecture modeling methods, automated approach can hardly provide plausible results while poses limitations on image number and shooting angle; interactive approach demands large amount of user sketches to form a complete geometry, which is inefficient. In order to effectively combine automatic analysis and user interaction thereby improving the user experience and modeling efficiency and robustness, this project studies integrating architecture priories into the analysis of image segments, automatic speculating and propagating of architecture geometry and initiative visualization of the uncertainties arisen therein to inspire users' correction and guidance. Three key scientific issues are to be addressed in this project: automatic geometry speculating and propagating, initiative heuristic visualization and user interaction law mining and utilization. This project proposes a computational framework that effectively fuses architecture images, architecture priories and user interactions to solve the architecture geometry, aims at achieving simple and efficient architecture modeling through the development of new interactive architecture modeling technology.

随着各种可拍照设备的普及，图片获取已非常容易，如何用图片快速重建现实世界三维模型仍然是计算机图形学和视觉领域的热点和难点研究问题。建筑作为城市场景中最重要的组成部分，其三维模型的构建受到学术界和产业界的广泛关注。现有基于图片的建筑建模中，自动化方法建模效果欠佳且对输入图片数量和拍摄角度要求较高；而交互式方法需要用户通过大量勾画形成完整几何结构，效率低。本项目主要研究如何结合建筑先验分析图片线段空间关系，自动地推测及扩展建筑几何结构，并主动呈现自动分析的不确定性以启发用户纠正和引导系统进行三维重建，实现自动分析与用户交互有机结合，从而改善用户体验，提高建模效率与鲁棒性。拟解决关键科学问题包括几何结构的自动推测与扩展、主动启发式呈现及用户交互规律挖掘与运用。项目拟提出有效融合建筑图片信息、建筑先验知识及用户交互求解几何结构的计算框架，旨在发展新的交互式建筑建模技术，实现简便高效的建筑建模。

随着各种颜色以及深度相机的普及，图片以及深度信息获取已非常容易，如何用图片或者深度图片快速重建现实世界三维模型仍然是计算机图形学和视觉领域的热点和难点研究问题。本项目提出使用大规模三维模型库（ShapeNet）来解决三维重建过程中的多个关键问题，包括基于三维模型形变子空间的图片深度信息恢复（发表于SIGGRAPH 2014），基于三维模型检索的室内场景重建（发表于EuroGraphics 2015），基于场景合成的室内场景重建（发表于SIGGRAPH Asia 2015）以及基于三维模型与图片联合嵌入的形状-图片检索（发表于SIGGRAPH Asia 2015）。同时，本项目提出通过检测形变事件来分析时序三维点云序列（发表于SIGGRPAH Asia 2013），以及通过介入式三维获取来精细重建带遮挡物体的三维结构（发表于CGF 2015）。