三维模型语义分析与检索研究

项目名称： 三维模型语义分析与检索研究

项目编号： No.60803073

项目类型： 青年科学基金项目

立项/批准年度： 2009

项目学科： 金属学与金属工艺

项目作者： 万丽莉

作者单位： 北京交通大学

项目金额： 20万元

中文摘要： 为了提高检索的准确度，研究三维模型检索的优化方法，提出了基于评价指标组合特征的三维模型检索方法、基于关键词和形状检索相结合的三维模型检索方法。为了实现模糊分类与多级决策，提出了基于凸包的二维形状分割算法，并结合视图特征和基于分割的结构特征，提出了基于模糊KNN和贝叶斯决策的三维模型分类方法。为了解决实验数据过少的问题，开展对三维模型数据收集方法的研究，提出基于Google 3D Warehouse扩充实验数据并实现半自动标注的方法、基于未定标序列的三维重建方法。为了检索和标注含有多个物体的三维模型，将多示例多标注学习框架引入三维模型语义分析领域，提出了基于包围盒相交检测、基于层次化聚类的三维模型多示例生成算法，以及基于多示例的三维模型检索与语义标注方法。为了检索非刚性物体，提出了基于自适应几何词汇选择的非刚体三维模型检索方法。此外，还形成了三维模型批量下载与格式转换工具、基于模糊分类的三维模型半自动标注系统、基于多示例的三维模型检索与语义标注系统等一系列辅助工具。研究成果可用于实现基于关键词的三维模型检索引擎，对高效获取和重用飞速增长的三维模型数据资源具有积极意义。

中文关键词： 三维模型；三维模型检索；三维模型语义标注；形状分类

英文摘要： In order to improve the effectiveness of 3D model retrieval, we investigated the optimization methods of 3D model retrieval, and then proposed an approach of feature combination by means of quantitative evaluation indicators, a 3D model retrieval method combining the keywords-based search method and the shape-based search method. In fuzzy classification and multiple decisions, we proposed a method for measuring concavities and segmenting a 2D shape by 2D convex hulls, and then presented a fuzzy KNN and Bayesian decision based classification method using view features and structural features. To increase the model number for experiments, we studied some methods to collect 3D models, and therefore presented a 3D model acquisition and auto-tagging method utilizing Google 3D Warehouse, and some approaches to reconstruct 3D models based on uncalibrated image sequences. To retrieve or label 3D models composed of multiple 3D objects, we leaded multi-instance multi-label learning (MIML) into 3D domain, and proposed 3D model's multi-instances generation algorithms by axis-aligned bounding boxes' intersecting criterion or hierarchical clustering criterion, also discussed two applications including 3D model retrieval and semantic annotation. For non-rigid model retrieval, we proposed a novel solution by adaptively selecting the set of geometry words for each query request. In addition, we developed a tool to download 3D models and convert their file formats, a 3D model semi-automatic annotation system, and multi-instance based 3D model retrieval and annotation system. The above researches can be used to establish the keywords-based 3D model retrieval engine, and have great significance to acquire and reuse rapidly increased 3D data resources.

英文关键词： 3D model; 3D model retrieval; 3D model semantic annotation; shape classification