objaverse-handling

Code for creating a dataset with 3D models from Objaverse. It allows to:

• download a given number of objects from chosen categories
• creating a subset of 3D models
• keeping track of ids and metadata

Usage

1. Git fork the repo
2. Create a python/conda environment with python>=3.7 (I personally use 3.9) and activate it.
3. Inside the project root directory, pip install the packafe locally with the commmand:

pip install -r requirements.txt

4. The file config.ini shows the paths used by default. Change any of these variables as necessary.

When downloading objects for the first time

One has to specify:

• the file from which to get the names of the categories (.csv format)
• the number of categories to download
• the number of objects to download from each category
• the file where to store the dictionary of uids
• the name of the worksheet in which will be stored the location of the downloaded models.

When downloading missing objects

One has to specify:

• the file with the dictionary of uids (that has been modified when curating the models)
• the file with the removed uids (created when curating the models)
• the file where to store the new dictionary of uids
• the worksheet where to store updated location of the downloaded models.

Note

In this version, the code takes objects iteratively, one after the other in the order they appear in the LVIS annotations dictionary. The selection is not randomized.