kimjson / kaist-3dml-assignments-1

Programming Assignments of CS492A: ML for 3D

KAIST CS492(A): Machine Learning for 3D Data (Spring 2022)
Programming Assignments

Instructor: Minhyuk Sung ([e-mail])
TA: Juil Koo ([e-mail])
Course webpage: https://mhsung.github.io/kaist-cs492a-spring-2022/
Last Updated: Mar 16, 2022.

Introduction

PointNet is a fundamental yet powerful neural network processing point cloud data. In the following programming assignments, we will learn how to use PointNet for different tasks including classification, auto-encoding, and segmentation.

Programming Assignment 1 (Classification / Auto-Encoding)

Due: Mar 21 (Mon) 23:59 KST

The goal of the first assignment is to run our code for classification and auto-encoding and see how it works. You just need to run our code and report the outputs! Our code includes a script automatically downloading the ModelNet dataset and a dataloader.

Run train_cls.py and train_ae.py, then you will see the test results after training:

• Classification

• Auto-Encoder

What to hand in

In a single document, write your name and student ID, add screenshots of the outputs (one for each task; thus two) like the ones above, and submit the document as a PDF file to KLMS.

If you want, you can implement the PointNet classification network and the autoencoder yourself! (e.g., using TensorFlow). In this case, submit your code (along with the snapshot document) as a ZIP file.

If you implement chamfer distance yourself, take the mean across both batches and points.
I.e., given P1 [B,N,3] and P2 [B,N,3], Loss = ChamferDistance(P1, P2) / (B*N), where B and N are the batch size and the number of points, respectively.

Grading

You'll get a zero score if you implement yourself but don't submit your code. Otherwise,

• 100% score:
  • If you submit on time, and
  • the document includes everything (your name, student ID, and two snapshots; one for classification and the other for auto-encoding),
  • and your results are on par with ours:

Classification	Auto-Encoding
test acc >= 85%	test loss <= 0.005

• 50% score:

  • If you submit on time, but
    • either one of the snapshots is missing, OR
    • one of the results is worse than the threshold above.

• 0% score: Otherwise.

Programming Assignment 2 (Segmentation)

Due: Mar 30 (Wed) 23:59 KST

The goal of the second assignment is to implement PointNet segmentation network based on our code. Our code includes script automatically downloading the ShapeNet dataset, a dataloader, main training/test loops, and the evaluation code computing mIoU. Implement the PointNet segmentation network in the PointNetPartSeg class of the model.py file. You can even use the code in the Internet to fill in this part!

Check out Figure 2 (not Figure 9) in the paper) for the details of the architecture, and feel free to leave out T-Nets

Run train_seg.py after filling in the PointNetPartSeg class of the model.py file. Then, the code will show the test mIoU and save some images (segmentation_samples.png) showing the part segmentation results:

What to hand in

Submit the followings to KLMS:

• A single PDF file including your name, student ID, a screenshot for mIoU, and the output image visualizing some segmentation results (like the ones above).
• A ZIP file including ALL the code (not just the part you implemented but all).

You can also implement the entire pipeline yourself if you want.

Grading

You'll get a zero score if you don't submit your code. Otherwise,

• 100% score:
  • If you submit on time, and
  • the document includes everything (your name, student ID, and a mIoU snapshot, a segmentation visualization image),
  • and your result is on par with ours:

Segmentation
mIoU >= 80%

• 50% score:

  • If you submit on time, but
    • either one of the results (a mIoU snapshot or a segmentation visualization image) is missing, OR
    • the result is worse than the threshold above.

• 0% score: Otherwise.