基于弱监督深度学习的三维模型分类方法研究

3D shape classification is an important method for geometric understanding of 3D shape, which is not only the basic of 3D structural analysis and generation, but also very useful to many applications, such as augmented reality, robot navigation, etc. Nowadays, a lot more 3D shapes are generated for multiple usages, which makes it possible, as well as crucial, to develop effective 3D shape classification methods. In order to classify 3D shape by a cheap, flexible and scalable way, this project combines the idea of weakly supervised learning and deep learning, which aims to realize the end-to-end classification under different supervision. First, it tactfully embeds the diverse property of 3D data representation into the co-training framework, which realizes co-training based 3D shape classification under incomplete supervision by multi-view and point representation. Second, it iteratively selects the high-quality data from the 3D shape set of the web to expand the training set, which realizes 3D shape classification under noise supervision by progressive sample filtering. Finally, an optimization framework is proposed by combining manifold and hierarchical relationship of the labels to recover the fine-grained tags from the coarse-grained annotation, which realizes 3D shape classification under inexact supervision by incorporating the manifold hypothesis and hierarchical constraint. The achievement of this project is to reduce the labeling quantity in 3D feature learning and promote the development of 3D shape in real applications.

三维模型分类是实现三维模型几何理解的重要手段，其不仅是三维模型结构分析和生成的基础，而且可直接用于增强现实、机器人导航等应用，而三维大数据时代的到临更增加了对三维模型分类技术的需求。为实现廉价、灵活和可扩展的三维模型分类，本项目拟结合采用弱监督学习和深度学习思想，拟在不同的弱监督条件下实现端到端的三维模型分类。首先，拟结合三维模型表示的多样性，采用基于分歧的半监督算法框架，实现基于视图投影和点云协同训练的部分标注三维模型分类；其次，拟以迭代方式选择互联网三维模型库中的高质量标注数据对训练标准数据集进行扩充，实现基于渐进样本过滤的噪音标注三维模型分类；最后，拟提出一种结合流形和标签层次关系的优化框架将粗糙标注恢复为细粒度标签，实现融合流形假设和标签层级约束的非精确标注三维模型分类。项目研究成果可显著降低分类所需的标注成本，进一步推动三维模型在实际应用领域的发展。

本项目以降低基于深度学习的三维模型分类的标注成本、提高三维模型分类的灵活性和可扩展性为目标，研究面向部分标注、噪音标注和非精确标注的三维模型分类方法，形成并建立完整的基于弱监督深度学习的三维模型分类技术和方法体系，从而推动三维模型在实际应用领域的发展。首先，针对面向部分标注的三维模型分类，依次提出了基于视图投影和点云协同训练的视图协同训练方法与基于Fixmatch的三维模型分类方法。其次，针对面向噪音标注的三维模型分类，依次提出了基于元学习损失矫正的三维模型分类和融合注意力机制和视图定位的三维模型分类方法。最后，针对面向非精确标注的三维模型分类，依次提出了融合主动学习、在线学习和互补标记的三维模型分类和基于无监督领域适应的三维模型分类方法。本项目发表论文9篇，其中SCI收录8篇，EI收录8篇，申请专利2项，参加国内外学术交流会议3次，并培养10名硕士，其中毕业硕士3名。