Machine Learning & Matching: High Volume Data Acquisition and Reconstruction Pipelines

In my contribution to this project, I developed an algorithm that computes 3D surface normals from 2D scans with sub-millimeter accuracy, capturing geometric discontinuities, textures and brushstroke orientations that prove more effective than traditional color and 3D shape for matching. I also developed a multifeature matching pipeline that uses logistic regression to find matches. The system out-performed state-of-the-art 3D matching algorithms, solving problems with scalability, efficiency, data validation and usability.  New matches were found in seconds (manual methods take weeks to years).