Error-Bounded Generative Models for 3D Scan Completion

用于 3D 扫描完成的误差有限生成模型

• 批准号: 10022640
• 金额: $ 3.77万
• 依托单位: VISIO IMPULSE LIMITED
• 依托单位国家: 英国
• 项目类别: Collaborative R&D
• 财政年份: 2022
• 资助国家: 英国
• 起止时间: 2022 至 无数据
• 项目状态: 已结题
• 来源: https://gtr.ukri.org/projects?ref=10022640
• 关键词: Error; Bounded; Generative; Models; 3D

Modern Machine Learning data analysis promise fast and flexible data processing, allowing reduced cost and lower barrier to entry for many 3D data acquisition tasks (e.g. laser scans). One problem regularly faced when trying to incorporate data pre-processing steps into downstream tasks such as creation of Digital Twins is to measure the accuracy and reliability of the system generated geometry.Having solid error bounds for the accuracy of the generated/processed 3D structure accuracy is the key enabler/barrier to adaptation of these techniques in markets such as Energy and Construction.The standard practice in data community is to measure performance on a separate validation dataset. This gives some indication into the performance of the ML models however, the overall question around the reliability of the geometric model remains. In addition, the best practice for selecting an unbiased and representative validation set is currently not well understood.Visio Impulse (https://www.visioimpulse.com) will work closely with the National Physical Laboratory, NPL, to explore various techniques for measuring accuracy and reliability of the ML generated 3D structure with the aim of providing certainty for geometry sensitive processes.The expertise and instrumentation available at NPL will be used to determine the correct approach to measuring geometric errors and the certainty bounds around the 3D structure modelling techniques, the dependencies and sensitivity to various elements within the model as well as the sensitivity to data acquisition approaches.This will allow the development of optimised (hybrid) strategy for 3D data acquisition and post processing and will be tested to show consistent performance, thus elevating the barriers to adoption of these techniques.

现代机器学习数据分析有望实现快速灵活的数据处理，从而降低许多3D数据采集任务（例如激光扫描）的成本和进入门槛。当试图将数据预处理步骤合并到下游任务（例如创建数字孪生模型）中时，经常面临的一个问题是测量系统生成的几何形状的准确性和可靠性。对于生成/处理的3D结构准确性的准确性，具有可靠的误差界限是关键的推动因素。在能源和建筑等市场中适应这些技术的障碍。数据社区的标准做法是在单独的验证数据集上衡量性能。这对ML模型的性能提供了一些指示，但是，围绕几何模型可靠性的总体问题仍然存在。此外，选择无偏和代表性验证集的最佳实践目前还不太清楚。NPL Impulse（https：www.visioimpulse.com）将与国家物理实验室（NPL）密切合作，探索用于测量ML生成的3D结构的准确性和可靠性的各种技术，旨在为几何敏感过程提供确定性。NPL提供的专业知识和仪器将用于确定测量几何误差的正确方法以及3D结构建模技术的确定性界限，这将允许开发用于3D数据采集和后处理的优化（混合）策略，并将进行测试以显示一致的性能，从而提高采用这些技术的障碍。