Assignment #2

Neeraj Basu (neerajb)

Zero late days used.

Late Days

Question 1.1

      src               tgt

Question 1.2

      src               tgt

Question 1.3

      src               tgt

Question 2.1

Question 2.2

Question 2.3

Question 2.4

F1 Score @0.05
Vox	27.475
Point	87.260
Mesh	80.247

Observing the numbers above, it's clear pointclouds yield the highest F1 score. This is due to the fact that pointclouds are not required to learn connectivity information like meshes. In addition, they are more flexible in the range of 3D points they can represent since they are not being discretized like voxels. Meshes performed slightly worse since they are required to learn connectivity and to respect connectivity using a smoothing loss. We also sampled points for the mesh which would lead to worse surface reconstruction. Finally, the voxel representation was the worst and this is due to the fact that our resolution was fairly small making good surface reconstruction difficult.

Question 2.5

F1 Score @0.05
Ico-Sphere 4	80.247
Ico-Sphere 5	79.611
Cow Obj	79.473

I played around with different initializations of the prediction mesh. I started by increase the subdivision level of the geodesic polyhedron from 4 to 5. It was no surprise this was able to deform smoothly to the surface of the chair. I then tried an object with higher frequency features, the cow from the first assignment. After ~3000 training iterations, the cow was also able to replicate the surface of the chair with a similar F1 score which I found surprising. While some of the meshes were not as smooth, with more training it's likely this could have improved.

Question 2.6

To help visualize what the net was doing during training, I visualized the same image in a batch during different portions of the training. With these visualizations, you can see how the original pointclouds and meshes are deformed to create the surface of the ground truth image.