April 23, Upsampling point clouds via deep learning

The goal of this session is to learn about state-of-the-art research in terms of deep learning for 3D point cloud processing. The session consists of a reading assignment of a recent research paper, and the learning check involves answering a few questions about the paper.

Reading assignment:

PU-GAN: a Point Cloud Upsampling Adversarial Network Links to an external site., Li et al., ICCV 2019 (webpage and code Links to an external site.)

Learning check:

Answer the following questions in  a few sentences and submit here:

1. Describe the problem statement the paper addresses in a few sentences.
2. Many components of their network architecture are built on MLPs (multi layer perceptrons). Describe what an MLP is in a few sentences (look it up on Wikipedia if you have not heard about it before).
3. Their point set generator shown in Fig. 2 consists of three main steps: feature extraction, feature expansion, and point set generation. Describe the purpose of each step in a few sentences. For each step, mention what its input and output is, and describe its size (dimensionality), that is, explain what r, N, C, C' means in the figure.
4. They use a loss function that consists of multiple components. Describe each component in a few sentences. Mention the motivation for using each component of the loss.
5. Write down one question that you have about the paper. I will collect all questions, try to answer them briefly, and post the answers to everybody.

Q&A

Here is a link to a summary of your questions and my answers from the learning check: https://docs.google.com/document/d/1egX51LlE29L5PfA9qSoq5nw56xR-suhDbyCYStLZkX8/edit?usp=sharing Links to an external site.