16-825 Assignment 2: R

Task 1.1 — voxel

Result:

Task 1.2 — point cloud

Result:

Task 1.3 — mesh

Result:

Task 2.1 — voxel pred

3 examples:

Task 2.2 — point cloud pred

3 examples:

Task 2.3 — mesh pred

3 examples:

Task 2.4 — Quantitative comparisions

Result:

Task 2.5 — Hyperparameter

Changed start mesh to a cow. My predictions are all cow flavored (aka predictions all have cow-like features/topologies).

Task 2.6 — Interpret your model

Result:

One paper I looked at is https://github.com/nywang16/Pixel2Mesh and tried to implement it for point clouds.

Task 3 — Exploring other architectures / datasets.

I chose to Extended dataset for training, by training points with all 3 classes

This result makes sense because the topology of a plane, car, and chair are very different. The resulting blob is due to the model needing to learn one latent mapping to 3 different topologies. The blob indicates that our model learned an average of the three topologies. I'm not sure if theres a classification associated with each of the models in the dataset or if we learn based on a classification (although i see no indication of it in the code), but including class label may help with this blob situation.