SBIR Phase I: Artificial Intelligence for Automated Custom Avatar Creation

这个小型企业创新研究(SBIR)第一阶段项目将为经验丰富的动画师创造一种快速制作高质量三维(3D)内容的方法，并为新手创造引人入胜的3D内容，而不需要多年的技术培训或强大但昂贵的软件。吸引人的计算机图形内容给教育、娱乐、医疗和虚拟环境带来了革命性的变化。该项目将使用人工智能(AI)通过缓解工作流程中的主要瓶颈来释放3D图形内容的全部潜力，3D图形内容的年市场价值为172.1亿美元。虽然美国目前有超过6.2万名动画师，但专门在3D动画工作室工作的不到1万人，而且他们中拥有重量画技能的比例更小。重量绘画是3D设计中的一项重要技术，它为角色增加了真实感，使他们在动画中能够流畅地移动。增加难度的是，权重绘制3D模型是一项单调乏味的任务，专家可能需要长达2天(约16个工作小时)才能手动完成一个模型。规模较小的动画商店往往根本没有完成这项任务的专业知识，也无法竞争更大、更有利可图的合同。此外，世界各地大学的研究人员和学生往往无法执行这种加权任务，这降低了他们在增强现实或虚拟现实中为医疗、体育和娱乐用途创建动画的能力。这个小型企业创新研究(SBIR)第一阶段项目将利用深度神经网络(DNN)从文本输入创建3D模型，并从行业标准的骨架系统和3D模型网格创建权重绘制的钻机。该技术将网格和骨架转换为机器学习(ML)代码可以处理的格式，引入了一种全新的数据结构。此外，该项目将探索COO(坐标列表)矩阵的改编，COO(坐标列表)矩阵是一种稀疏矩阵，具有神经网络的有效性能，但在应用于坐标排序不确定的3D空间中的机器学习任务时面临挑战。最困难的问题，如权重绘制和建模，一直受到四个具体限制的阻碍：1.缺乏地面真实数据，2.有限的训练数据，3.缺乏先验最大似然架构，4.缺乏对非高斯数据的稳健性和特异性。该项目将通过建立纳入和转换用于DNN分析的非高斯数据的方法来进入这些领域，并将创建一个全面的数据训练集，同时基于规范的图灵测试建立特定领域的地面真相。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

This Small Business Innovation Research (SBIR) Phase I project will create a way for experienced animators to rapidly make high quality three-dimensional (3D) content, and for novices to create engaging 3D content without the need for years of technical training or powerful but expensive software. Engaging computer graphic content has revolutionized education, entertainment, medical, and virtual environments. This project will use artificial intelligence (AI) to unlock the full potential of 3D graphical content, a $17.21 billion annual market, by alleviating major bottlenecks in the workflow. While there are more than 62,000 animators currently employed in the United States, fewer than 10,000 work specifically at 3D animation studios, and a smaller proportion of them possess the skills for weight painting. Weight painting is a vital technique in 3D design that adds realism to characters, enabling them to move smoothly during animation. Compounding the difficulty, weight painting 3D models is a tedious task that can take an expert up to 2 days (or around 16 work hours) to manually complete one model. Smaller animation shops often do not have the expertise to perform this task at all and are unable to compete for bigger, more lucrative contracts. Furthermore, researchers and students at universities around the world are often unable to perform this weighting task, which reduces their ability to create animations for medical, athletic, and entertainment uses in augmented reality or virtual reality. This Small Business Innovation Research (SBIR) Phase I project will utilize deep neural networks (DNN) to create 3D models from text input as well as a weight-painted rig from an industry-standard skeleton system and a 3D model mesh. The technology converts the mesh and skeleton into a format that can be processed by machine learning (ML) code, introducing a brand-new data structure. Additionally, the project will explore an adaptation of the COO (Coordinate List) matrix, a sparse matrix that performs effectively with neural networks but faces challenges when applied to machine learning tasks in 3D space where coordinate ordering is uncertain. The most difficult issues, such as weight painting and modeling, have been hampered by four specific limitations: 1. Lack of ground-truth, 2. Limited training data, 3. Lack of a-priori ML architecture, and 4. Lack of robustness and specificity for non-gaussian data. This project will make inroads into each of these areas by establishing a methodology for incorporating and transforming non-gaussian data for DNN analysis and will create a comprehensive data training set while establishing a domain specific ground-truth based on the canonical Turing test.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.