April 28, CNNs on meshes

Slide material:

23 Data-driven shape modeling - CNNs on meshes.pdf Download 23 Data-driven shape modeling - CNNs on meshes.pdf

Panopto recording:

Learning check:

The learning check involves reading the following paper and answering a few questions about it.

Paper reading assignment: "Spherical CNNs on unstructured grids Links to an external site.", Jiang et al., ICLR 2019 (code on github Links to an external site.).

Questions:

• Describe the problem statement addressed by the paper.
• The description of Equation 1 says "The 3 × 3 kernel  [...] can be written as a linear combination of basis kernels which can be viewed as delta functions at constant offsets".  Try to explain this statement in your own words in 2-3 sentences.
• They say (in the paragraph before Eq. 4) "we replace the cross-correlation linear operators [...] with differential operators of varying orders".  Explain (1) the motivation to do this, and (2) why this works to construct neural networks in a few sentences.
• In Section 4.2 they discuss object classification using their approach. This involves preprocessing the original meshes  into spherical signals. Explain why they do this, instead of just applying their approach directly to the meshes.
• Write down one question you have about the paper.

Q&A

Here is the feedback on your questions from the learning check:

https://docs.google.com/document/d/16UQ62e4nPwVsad91JqVEzQ_5xCD_VIeGQOmHahgZYuU Links to an external site.