项目名称: 三维场景理解中的高阶能量优化理论与方法研究
项目编号: No.61333015
项目类型: 重点项目
立项/批准年度: 2014
项目学科: 自动化技术、计算机技术
项目作者: 胡占义
作者单位: 中国科学院自动化研究所
项目金额: 300万元
中文摘要: 本项目旨在研究探索如何通过在随机场中引入合适的高阶能量模型来有效表达关于场景的定性和统计先验知识,将二维图像与三维点云统一在同一能量优化框架下,同时提高三维场景理解能力和三维物体重建精度的理论和方法。本项目的三维场景理解包括物体分割与识别、物体定位、物体三维形状完整化等内容。设计合理的高阶能量模型和建立有效的高阶能量优化算法是本项目的关键问题。 本项目的主要研究内容包括以下四方面:(1):三维场景理解中的高阶能量有效表达问题;(2):含高阶能量项的快速能量优化方法;(3):高阶能量模型与层次化网络模型之间的关系, 如深度神经网络;(4):基于GPU+CPU的混合快速能量优化算法。 本项目的研究可望建立一套比较系统的针对三维场景理解的高阶能量模型设计和快速优化的理论和方法,形成相关的算法模块,并能在一些特定领域得到初步应用。本项目的预期成果可望进一步丰富和推动三维场景理解的研究与进展。
中文关键词: 计算机视觉;三维场景理解;三维形状完整化;高阶能量(HOE)表达;高阶能量优化
英文摘要: This project is intended to explore appropriate higher order models in random fields to better express qualitative and statistical knowledge about the scene, and effective inference methods to simultaneously improve the ability of 3D scene understanding and to complete 3D shapes under a unified framework by integrating 2D images and the corresponding 3D point clouds. In this project, 3D understanding is meant to segment and recognize objects, to determine object’s pose, and to complete incompletely reconstructed object shapes etc. Here design appropriate higher order models and establish effective inference methods for 3D scene understanding are the core issues. The project’s main contents include the following 4 parts: (1): Design appropriate higher order models for 3D scene understanding; (2): establish effective inference methods for higher order energy minimization; (3): investigate the relationship between higher order model and hierarchical networks,such as Deep Neural Networks ( DNN) ;(4): GPU+CPU mixed fast implementation schemes. The expected outcome is a comprehensive theory and method for higher order models design and effective inference in 3D scene understanding, as well as some fast algorithms. In addition, the project’s results are expected to find real applications in some specific domain. Finally, this study will enrich 3D scene understanding research and promote its further development.
英文关键词: Computer vision ;3D scene understanding;3D shape completion;HOE representation;HOE minimization