| 摘要: |
| 目的 通过对近年来基于深度学习的三维形状智能生成方法进行梳理和分析，总结该方向存在的问题与挑战，预测未来发展趋势，为相关研究者提供参考。方法 首先从数据表示角度分析三维形状智能生成中常用的体素、点云、网格、隐函数等表示方法的优缺点并总结了常用的数据集。其次从三维形状生成内容的可控性出发，综述了目前研究中整体生成、结构感知生成和交互式生成方法。最后从生成结果的相似性、多样性和结构一致性方面总结了在三维形状智能生成中常用的评价指标。结论 三维形状智能生成领域虽然已经取得明显发展，但是在数据集的规模、生成方法的有效性和评价指标的全面性方面仍然面临较大挑战。 |
| 关键词: 三维形状生成 深度学习 形状表示 智能生成 形状评价 |
| DOI：10.19554/j.cnki.1001-3563.2021.22.011 |
| 分类号:TB472 |
| 基金项目:国家自然科学基金(61872137);湖南省科技创新计划(2021RC3064);湖南省自然科学基金(2020JJ4009) |
|
| Summary of Approaches of Intelligent 3D Shape Generation |
|
XIAO Yi, QIAO Jin-hao, JI Tie
|
|
(Hunan University, Changsha 410082, China)
|
| Abstract: |
| By investigating and analyzing the deep learning based approaches in 3D shape intelligent generation in recent years, this paper summarizes the problems and challenges of this field and predict the tendency of the future work, to provide a technical reference for researchers in related fields. Firstly, the advantages and disadvantages of the commonly used 3D shape representation methods and datasets are analyzed, such as voxels, point clouds, meshes, implicit functions, etc. Secondly, from the perspective of controllability of 3D shape generation objects, holistic generative methods, structure-aware generation methods and interactive generation methods are reviewed. After that, the evaluation metrics commonly used in 3D shape intelligent generation, such as similarity, diversity and structural consistency of the generation results are summarized. Although the field of 3D shape intelligent generation has made significant development, it still faces great challenges in terms of the size of data sets, the effectiveness of generation methods and the comprehensiveness of evaluation metrics. |
| Key words: 3D shape generation deep learning shape representation intelligent generation shape evaluation |
