MattisKr/co3d-sam3d-affordance

⚠️ Slimmed to meshes-only for human affordance labeling. Heavy latent/feature files were removed to keep this lean; mesh.glb, vertex_positions.pt, vertex_normals.pt, the GEAL vertex_pseudolabels_*.pt, and the manifests are kept. See ANNOTATION_SPEC.md before labeling.

CO3D → SAM3D → GEAL Affordance Dataset

Per-vertex affordance pseudolabels on SAM 3D Objects reconstructions of real CO3D images, for verb-conditioned 3D affordance prediction. Single real image → SAM3D mesh → frozen GEAL teacher labels per-vertex affordance for each verb on the same geometry (labels already in mesh-vertex order, no cross-modal alignment).

Layout

• reconstructions/<object>/: mesh.glb, vertex_pseudolabels_<verb>.pt (V,) GEAL scores in [0,1], vertex_semantics.pt (per-vertex CLIP, fp16), dino_cls/ss_dino_cls/dino_patches/ss_dino_patches, slat_*.pt, global_latent.pt, gaussian.ply, vertex_normals.pt, meta.json.
• manifest*.jsonl: one row per (object, verb).

Categories / verbs

10 CO3D categories → GEAL classes (bottle, bowl, cup, chair, handbag, keyboard, laptop, microwave, tv, vase). Verbs: contain, grasp, pour, sit, move, lift, press, display, open.

License & attribution

Derived from CO3D (Common Objects in 3D, Meta AI; CC-BY-NC 4.0) → this dataset is CC-BY-NC 4.0 (non-commercial). Reconstructions via SAM 3D Objects (Meta); affordance labels via GEAL (CVPR 2025). Please cite CO3D, SAM3D, and GEAL. Raw CO3D imagery is not redistributed here.