We document here the usage of our 2D and 3D dataloaders. Our 2D dataloaders our wrappers over the WebDataset API.
For 3D, we provide a “zipped” dataloader that effectively enables you to sample 1000 average style augmentations per model. These augmentations are defined by embedded metadata, and consistent with the 3D pointclouds and renders with also provide.
We also provide a simple HDF5 dataloader in our repository, enabling you to quickly iterate over pre-sampled pointclouds.
If you are participating in the 3DCoMPaT++ Challenge, note that only pointclouds are available for the test set.