CPS: Medium: Collaborative Research: Srch3D: Efficient 3D Model Search via Online Manufacturing-specific Object Recognition and Automated Deep Learning-Based Design Classification

点击翻译按钮获取中文摘要

Rapid growth in additive manufacturing (AM) has improved the accessibility, customizability and affordability of making products using personal printers. Designs can be developed by consumers, if they have enough knowledge of mechanical design and 3D modeling, or they can be obtained from third parties. However, the process of translating a design to a program that can be successfully executed by a 3D printer often requires specialized domain knowledge that many end-users currently lack. In the meantime, lots of objects, which may be very similar or identical to what the non-technical user aims to design and print, have been produced by experts in industry and, hence, millions of proven part designs already exist. This research aims to fill the above-mentioned gap by developing a theoretically sound and practically deployable, domain-specific online search engine, called Srch3D, for 3D models. Srch3D will provide the non-technical end-users with a user-friendly solution to efficiently search for their components in a large repository of existing proven part designs. The outcomes of this project will include algorithms for advanced 3D model analysis, indexing and search algorithms that can identify designs of interest within a large number of proven design files accurately in runtime. The research will involve development of algorithms for automated design search via 3D object detection with adaptive resolutions. They will build on top of state-of-the-art computer vision techniques, namely histogram of gradients (HOG), and extend them to three-dimensional spaces for the manufacturing design files. Additionally, the project will research algorithms for runtime 3D object classification and labeling via data-driven modeling. The solutions will use deep neural networks to search and identify objects of interest from a large design repository. The use of relatively high-level data-driven models, along with the detailed HOG-based solutions, will enable our online 3D model search engine to accept a different variety of input object formats from the users, such as sketches or photos of the objects of interest, their (partial) G-Code, computer-aided design design files, or English descriptions and keywords. The framework will be accessible via a public cloud-based 3D model search service. In the vein of google.com and virustotal.com for document and malware search, respectively, the framework will realize the aforementioned modules as a cloud-based search engine service that allows anyone to search for their design of interest using different input formats.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.