synthetic-cad-view

Synthetic CADView / CAD-pose grounding data from Vitrus.

Each training example pairs a rendered tabletop scene with CAD reference imagery, a 32-view orientation bank for the queried part, and exact geometric supervision from a manifest. The dataset is designed for CAD-conditioned 3D pose grounding and render-and-compare orientation prediction.

Source Layout

Original Vitrus asset layout:

gs://vitrus-assets/cad_pose_grounding/v2/scenes
gs://vitrus-assets/cad_pose_grounding/v2/cad_refs
gs://vitrus-assets/cad_pose_grounding/v2/atlas/n32
gs://vitrus-assets/cad_pose_grounding/v2/symmetry_groups.json

Contents

• scenes.tar: rendered scenes and per-scene manifests, archived from scenes/.
• cad_refs.tar: canonical CAD reference renders and metadata, archived from cad_refs/.
• atlas/n32/: 32-view CAD orientation bank grids, stored directly as PNG files.
• symmetry_groups.json: per-part proper-rotation symmetry groups used for orbit-aware supervision and evaluation.
• machined_assets.json: asset manifest used by the CAD-pose generation pipeline.

Dataset Stats

• 30,226 scene folders in the source bucket.
• 6,321 CAD reference folders in the source bucket.
• 6,320 atlas images in atlas/n32.
• 6,320 symmetry groups, including 3,034 non-trivial orientation orbits.

Related Model

The paired Vitrus Prismatic model is published as vitrus/Qwen3.5-4b-prism-3D.

License

Apache-2.0 for this synthetic release package.

Archive Extraction

tar -xf scenes.tar
tar -xf cad_refs.tar

After extraction, the local layout matches the original source prefixes: scenes/, cad_refs/, atlas/n32/, symmetry_groups.json, and machined_assets.json.