Tutorial: Toilet Reconstruction with Models that Learn Geometry (Mamba Family)

5 models following-up from literature review

In my last post, “Literature Review: The Best ML Models for Learning Geometry”, I covered a number of publications for various models that can learn and reconstruct geometries. As follow-up, I wanted to share a code example that applies several of these to a common modeling task. I’ll share the code, data, and an overview blog to go through training these models and looking at the results.

The notebook is a small reconstruction tutorial built around one shape category: toilet from ModelNet10. Instead of asking the models to classify the object, the notebook asks each model to compress a toilet point cloud into a latent representation and then reconstruct the same point cloud. This makes the comparison more visual than a classification score. You can look at the generated point sets and ask whether the model has learned the bowl, tank, seat area, base, and overall proportions.