面向三维模型草图检索的异构数据匹配关键技术研究

基于草图的三维模型检索是数字3D产业中的一个重要核心需求，是三维建模、VR/AR、自动驾驶仿真、CAD设计等众多高层应用中基础而关键的一环。然而，草图和三维模型数据的复杂性、多样性以及两者的数据异构性，成为严重影响检索精度和制约产业发展的瓶颈。现有方法主要采用标准的深度卷积网络对草图和三维模型的匹配问题进行建模，缺乏对检索中数据特性和难点的针对性研究。本课题将深入研究基于草图的三维模型检索中的异构数据建模和匹配问题，旨在显著提升检索的精度和鲁棒性。围绕检索中的各项核心技术问题，课题在草图质量评估和噪声处理、匹配任务驱动的模型视角选取、草图与模型跨模态匹配精度提升等三个方面展开探索。在解决上述科学问题的基础上，建立统一的特征学习和检索框架，进而获得准确鲁棒的检索性能和自然直观的检索体验。本课题的开展将有效促进三维模型检索领域的应用和发展，为3D产业的智能化发展提供理论和应用基础。

Sketch-based 3D model retrieval is an important core requirement in the digital 3D industry. It is also a fundamental and critical module in many high-level applications such as 3D modeling, VR/AR, automatic driving simulation and CAD design. However, the complexity and diversity of sketch and 3D model data, as well as the heterogeneity between them, have become a bottleneck that seriously affects the accuracy of retrieval and restricts the development of the industry. Existing methods mainly use standard deep convolutional networks to model the matching problem between sketches and 3D models, and lack of targeted research on data characteristics and difficulties in retrieval. This project will deeply study the modeling and matching of heterogeneous data in sketch-based 3D model retrieval, aiming to significantly improve the accuracy and robustness of retrieval. Focusing on the core technical issues in retrieval, the project explores three aspects including quality assessment of sketches and noise processing, matching task driven model viewpoint selection, and improvement of cross-modal matching accuracy between sketches and models. On the basis of solving the above scientific problems, a unified feature learning and retrieval framework is established to obtain accurate and robust retrieval performance, and natural and intuitive retrieval experience. The development of this project will effectively promote the application and development of the field of 3D model retrieval, and provide a theoretical and application basis for the intelligentization of the 3D industry.