Hugging Face's logo

Datasets:
AuWang
/
PartNeXt_mesh

License:
Dataset card Files
xet
 Community
2

Annotation/mesh count mismatch: 23,519 annotated shapes but only 23,229 .glb files (290 shapes have no mesh)

#2
by faris-abuali - opened
Discussion
faris-abuali

Hi, and thank you for releasing PartNeXt — it's been very useful for our work.

While building a pipeline on top of PartNeXt, I found a count mismatch between the annotations and the meshes: 290 shapes are annotated but have no corresponding .glb in PartNeXt_mesh.

The numbers

Source Count
Shapes per the AuWang/PartNeXt dataset card ("Number of rows") 23,519
Annotation set in AuWang/PartNeXt_raw (matches the card, and the set I downloaded) 23,519
.glb files in AuWang/PartNeXt_mesh 23,229
Difference 290

How to reproduce 

from huggingface_hub import HfApi
api = HfApi()
glbs = [f for f in api.list_repo_files("AuWang/PartNeXt_mesh", repo_type="dataset")
        if f.endswith(".glb")]
print(len(glbs))   # -> 23229, but the dataset has 23519 annotated shapes

I took the model_ids of the 290 shapes that have an annotation but no .glb, and confirmed that 0 of the 290 appear anywhere in PartNeXt_mesh — so this isn't a path/naming mismatch; the meshes are genuinely absent. The 290 are spread broadly across many categories (Chair, Table, Bookcase, Knife, Bucket, Mug, Glasses, Lamp, Pen, Sofa, Bed, …), not concentrated in one.

Sample of affected model_ids (annotated, but no .glb) 

32cfa73193de47c2b70c374ca170a8f8   (Bookcase)
b5e8f0d9d15a471c8047150a1bda61fd   (Skateboard)
d9c65fcd16314ee997987c974f8ace04   (Guitar)
febeceb3adaa45109f2c0be1d2c4e1ca   (Chair)
3a40214e5f654075aba1716d63ba323c   (Knife)
42b08255e31941d48d4d6444070c0ed8   (Chair)
8374abfafd5c477397d89cccbdc8407a   (Knife)
b073d50b0fb94025a08d7841433a059c   (Monitor)
b9c8a4e2e12d477d82b5c0f645a654ad   (Bookcase)
c8afa62b31674974933c837207e63354   (Table)

(Happy to share the complete list of all 290 model_ids if that helps.)

Questions

  1. Were these 290 meshes intentionally withheld, or is this an oversight in the PartNeXt_mesh upload?
  2. If it's an export/upload issue — would it be possible to add the missing 290 .glb files?
  3. If they were dropped deliberately (e.g. failed a QC step), could the dataset card / annotation set (currently 23,519) be reconciled to the 23,229 shapes that actually have meshes, so downstream users don't hit this gap?

Thanks again for the dataset, and for any clarification!

Sign up or log in to comment