ComAsset

ComAsset is the dataset of paper "Beyond the Contact: Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models".

The dataset consists of total 83 object meshes, collected from SketchFab.

All of the meshes are converted to .obj format with image texture files. We manually canonicalize the objects in terms of location, orientation and scale.

The format of the dataset is as follows:

ComAsset
├── data
│   ├── accordion # object category
│   │   └── wx75e99elm1yhyfxz1efg60luadp95sl # object id
│   │       ├── images # folder for texture files
│   │       ├── model.obj
│   │       └── model.mtl
│   ├── axe
│   ├── ...
│   └── watering can
└── categories.json

In categories.json, you can check the existing object categories, along with the original data URL and the license information.

License

ComAsset is licensed under the ODC-By v1.0 license. This license applies to the dataset as a whole, and users must also comply with the licenses of individual content. The license of each content is specified in categories.json.

Loading Dataset

from datasets import load_dataset
from huggingface_hub import snapshot_download
import trimesh

snapshot_dir = snapshot_download(repo_id="SShowbiz/ComAsset", repo_type="dataset")
comasset = load_dataset("SShowbiz/ComAsset", data_files={"metadata": "**/metadata.json"})

with open(os.path.join(snapshot_dir, "categories.json"), "r") as json_file: objects = json.load(json_file)
categories = [object_metadata["category"] for object_metadata in objects]

category, *_ = categories # first category
object_metadata, *_ = comasset['metadata'].filter(lambda example: example['category'] == category) # first object

obj_path = os.path.join(snapshot_dir, object_metadata["obj_file"]) 
mesh = trimesh.load(obj_path)

Citation

To cite ComA, please use the following BibTeX entry:

@inproceedings{ComA,
  title="Beyond the Contact: Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models",
  author="Kim, Hyeonwoo and Han, Sookwan and Kwon, Patrick and Joo, Hanbyul",
  booktitle=ECCV,
  year={2024}
}