daloibl/oscar-plus-galleries

OSCAR+ preprocessed galleries

Preprocessed CAD galleries for the OSCAR+ code release: rendered reference views, partial point clouds, view/caption/shape embedding caches, and the dGeDi geometric descriptor galleries. Building these from scratch takes several GPU-days, so they are published here to make the evaluation reproducible in reasonable time.

Code, evaluation scripts and the frozen results: https://github.com/ThoLoib/OSCAR-plus

What is in here, and what is not

Included (all of it generated by this project):

Path	Content
object_images/<gallery>/<object>/	42 rendered views per object (<id>_<v>.png), the background view, the camera matrices, and the per-view partial point clouds (*_partial.npz)
object_images/<gallery>/.*cache*.pt	embedding caches: DINOv2 and SigLIP over the views, ULIP-2 and Uni3D over the partial clouds
object_images/<gallery>/precompute_manifest.json	which encoder pass produced which cache
object_database/<gallery>/descriptions_attributes.json	the LLaVA captions, one per rendered view
caches/dgedi/bop/, caches/dgedi/shrec/	dGeDi descriptors, 1 316 and 3 308 objects

Not included: the CAD models and the benchmark images themselves. Those belong to their original datasets (SHREC'18, MI3DOR, YCB-V, T-LESS, LM-O, ITODD, Google Scanned Objects, HouseCat6D); download them from their sources and accept their licences there. The release README lists every source.

Galleries

Gallery	Objects	Used by
shrec18	3 308	Stage 1 (retrieval design)
MI3DOR	3 848	Stage 2 (image-only retrieval)
ycbv, tless, lmo	21 / 30 / 8	Stage 3 targets, Stage 4, Stage 5
gso, housecat6d, itodd	1 030 / 199 / 28	the proxy gallery of Stage 3
demo_gallery	2	the two-object demo of the README

Use

Point the release's config/paths.yaml at wherever you put these folders, or place them at the repository root so the defaults find them:

pip install huggingface_hub

# everything (~69 GB)
hf download daloibl/oscar-plus-galleries --repo-type=dataset --local-dir .

# or only what one stage needs, for example Stage 3/4/5:
hf download daloibl/oscar-plus-galleries --repo-type=dataset --local-dir . \
    --include "object_images/{ycbv,tless,lmo,gso,housecat6d,itodd}/*" \
               "object_database/*" "caches/dgedi/bop/*"

The embedding caches are content-fingerprinted over the rendered views, so the views have to be present as well: a cache without its renderings is ignored and the encoders run again.

Provenance

Produced with the release code at the commit recorded in each precompute_manifest.json, on an RTX 4090 with PyTorch 2.10 / CUDA 12.8. Renderings come from Blender 3.4.1 (42 icosphere viewpoints, FPS-ordered, so the first N views of any object are a valid N-view subset). Two galleries were built with non-default settings, which is why their numbers are what they are: MI3DOR's partial clouds use a hidden-point-removal parameter of 3.2 and no jitter, everything else uses 2.8 and 0.001.

Licence

The generated artefacts in this repository are released under CC BY 4.0. They are derived from the datasets listed above, whose own terms continue to apply to anything derived from them; ITODD in particular is restricted to non-commercial research.