Now, I'm exploring to download the dataset of Objaverse-XL.
The official Huggingface website is here.
- Clone the objaverse-xl repository.
git clone https://huggingface.co/datasets/allenai/objaverse-xl- Before using the test_api.py, you need to install the "pyarrow" or "fastparquet". I attempt the fastparquet and it works.
pip install fastparquet- Use the script "download_script_test.py" for testing the downloading and get some examples of smithsonian subdataset.
python download_script_test.pyThen, you will get some ".glb" files.
- I use the check_glb.py file to check the status of the ".glb" file.
python check_glb.py- Sample the point clouds and visualization
a) Sampling
The sampled point clouds can be obtained through sampling vertices or meshes (using open3d). (pip install open3d)
The trimesh is preferred. The code is referred by the openshape demo, from here.
python point_cloud_sampling.pyNote, I tried the smithsonian sub-dataset, but the first five files (.glb) seem not to be read by open3d, and similar errors are reported by trimesh. The ".glb" contains the color inforamtion.
I turn to try the thingiverse sub-dataset. The format is ".stl", which doesn't have any color information. I'm not sure about the effect of the color on rendering and point cloud-image-text contrastive learning.
b) Visualization (TODO)