Artifacts when reproducing SDF generation for training data (following AlignSDF)

[taeyunwoo]

Hi, thank you for releasing this great work and the repository.

I would like to clarify an issue regarding SDF generation.
The training SDFs are not provided, so I am currently generating them myself.
As a sanity check, I am first following the same SDF generation procedure used for the test dataset (AlignSDF), for which SDFs are provided.

However, when reproducing this process, I consistently observe strong artifacts after SDF-to-mesh reconstruction via marching cubes (e.g., fragmented components and floating surfaces, see attached images).

I understand that releasing the full SDF generation code may be difficult.
In that case, could you please provide guidance on the expected SDF data representation?

In particular:

• What is the expected hand/object mesh format for SDF generation (OBJ conventions, face indexing, watertight/manifold requirements)?
  • NOTE: I used watertight mesh faces to both hand and object.
• What dtype/precision is used for vertices and faces, and are there any assumptions on numerical accuracy?
• Are there any normalization or preprocessing steps applied (e.g., centering, scaling, unit conventions, or mesh cleanup)?

Any clarification on how the hand/object SDFs are generated or stored would be extremely helpful.
Thank you very much for your time.

[Left: provided object SDF to mesh / Right: generated (by me) object SDF to mesh / using same marching cubes algorithm]

[taeyunwoo]

Also, I have tried mesh saving code frag in AlignSDF (code), but it does not work.