ASSIGNMENT 7

In this assignment, we will load and visualize some 3D models as datasets in PyTorch3D.

The goals of this practice are the following:

  • Learn about the ShapeNet dataset
  • Understand the challenges in building a dataset useful for machine learning
  • Experiment with PyTorch3D DataLoaders
  • Visualize 3D models using Plotly

Instructions:

If you’re using Google Colab, you just need to have a google account and an associated Google Drive. Make a copy of the notebook located below and modify it as requested.

In case you’re choosing to work locally in your machine you must set Anaconda or a venv virtual environment, and install the necessary libraries. Create a folder in your Google Drive or in your machine’s workspace. Copy to your drive folder or download the following notebook:

Assignment 7 Notebook Open In Colab

  1. You’ll need to download a small subset of the ShapeNet dataset to complete this task. Download it here
  2. Follow the instructions in the notebook for completing the assignment.
  3. If you want, you can build auxiliary .py scripts and call them from your notebook, for organizational purposes.

Submission

The assignment is due on May 17th, 2023 at 11:59pm (GMT-3).

Students should send their assignments before the due date to hallpaz@impa.br with a copy to lvelho@impa.br. Late delivers will be consider subject to a lower score.

The submission email should be sent with the subject “Assignment 7 - [first-name] - [last-name]”. The assignment can be structured and sent in two ways:

If your whole solution is implemented in the same notebook as the one provided for the assignment, then you can send just the .ipynb file as the solution. If parts of your implementation were done in auxiliary .py scripts, then you must send both the final notebook and the scripts inside a .zip file. The organization of the code will also be considered in the evaluation.

References and other useful contents:

  1. https://shapenet.org
  2. ShapeNet: An Information-Rich 3D Model Repository
  3. PyTorch3D source code for Meshes data structure.
  4. ImageNet
  5. WordNet
  6. Objaverse
  7. Paul Bourke sobre arquivos .obj
  8. Synthetic Data Generation for Machine Learning
  9. Some parametrics surfaces for PyTorch3D (Github code)