mesh2sdf preprocessing for training vs inference

[zih-an]

I have downloaded and inspected the dataset on HuggingFace, and noticed that many meshes are not watertight and do not have preprocessed pc_normal. By reading the loop_set_256.py code, I also noticed that the training meshes are not preprocessed with mesh2sdf, whereas the inference procedure samples points from meshes that have been preprocessed with mesh2sdf.

I believe the main purpose of mesh2sdf is to ensure watertightness, which primarily affects the consistency of normal directions for sampled points. Could you clarify why this preprocessing step is considered necessary for inference but not for training?

[mush-room]

I have the same question. I'm looking forward to authors' reply.