April 21, Deep learning on point clouds

Slides:

22 Data-driven shape modeling - deep learning on point clouds.pdf Download 22 Data-driven shape modeling - deep learning on point clouds.pdf

Panopto recordings:

1. Deep learning on 3D shapes, introduction Links to an external site.
2. Convolutions on 3D shapes Links to an external site.
3. Convolutions on 3D point clouds Links to an external site.

Learning check:

The learning check is about the $\chi$-convolution approach for 3D point clouds that was presented in PointCNN: Convolution On X -Transformed Points Links to an external site.by Li et al., as discussed in the course material. By reading up the details in the paper, answer the following questions in a few sentences and enter the text here as your submission:

• Describe the strategies proposed in the paper to build hierarchical convolutions, that is, networks with multiple convolutional layers at lower resolutions.
• Describe how the approach is applied for shape segmentation tasks. How is the network architecture for segmentation different from the architecture for classification?
• Describe in words the difference between the methods in each column in the ablation test in Table 5 in the paper. Explain the motivation for the methods "w/o X-W" and "w/o X-D" (last two columns). In short, what do these results indicate?

Links to an external site. Links to an external site.