We provide instructions on how we preprocessed data for training IGR.
For the "bob and spot" shape space, we adapted the preprocessing script provided for the dfaust to generate surface points and normals for bob and spot, which we downloaded from https://www.cs.cmu.edu/~kmcrane/Projects/ModelRepository/.
For the ShapeNetV1 objects, we want to avoid points from internal surfaces. We thus adapted the rendering-based SampleVisibleMeshSurface program from DeepSDF to output surface normals for sampling 250000 points and storing points as well as normals as npy files.
We then created a data loader based on the one for dfaust to load the preprocessed data for training IGR.
We provide the training configurations used in our experiments in IGR_data/train_configs and the splits containing the object ids we used in IGR_data/splits.