Learning on 3D geometries - September 2021

Length: 12 hours - 2 cfu

Abstract

The course provides some basics for the application of Machine learning techniques to 3D shapes, combining aspects of Geometry Processing for the representation of 2-manifolds or volumetric structures and aspects of Machine Learning carried out on these structures.

Syllabus:

1st Lecture (3 hours)   14:30-17:30  13/09/2021

• Introduction to traditional ML,
  mage-oriented ML applied to computer vision (1.5 h - Melzi)
• Introduction to  geometry processing.
  Traditional digital representations for 3D models (1.5 h - Tarini)

2nd Lecture (3 hours)  14:30-17:30  15/09/2021

• Shape matching problem, ICP, non-rigid matching, local descriptors (1.5 h - Melzi)
• Spectral geometry processing (1.5 h - Melzi)

3rd Lecture (3 hours)  14:30-17:30  16/09/2021

• Convolutions on 3D data, multi-view CNN,  volumetric voxel-based structures (1 h - Melzi)
• Point-based convolution:
  spectral convolution, diffusion-based convolution, geodesic convolution (1.5 h - Melzi)
• Edge-based convolution Mesh CNN (0.5 h - Melzi)

4th Lecture (3 hours)  14:30-17:30  22/09/2021

• Functional Maps  and Data-driven approaches for Functional Maps (1.5 h - Melzi)
• PointNet and demo on region detection on human shapes (1 h - Melzi)
• Conclusion (0.5 h - Tarini)

Lecturers:

Prof. Marco Tarini (Associate Professor, Università degli Studi di Milano) – 2 hours
Dr. Simone Melzi (Researcher, Univerisità la Sapienza di Roma) – 10 hours

Assessor:

Prof. Marco Tarini (Associate Professor, Università degli Studi di Milano)
Dr. Simone Melzi (Researcher, Univerisità la Sapienza di Roma)