Preprocessing scripts for CLEVR dataset.

[Cuberick-Orion]

Hi,

First of all, amazing work!

I am wondering if you could share some light about the preprocessing of the CLEVR dataset -- specifically, scripts that correspond to Appendix B.1 in the paper.

My problem at hand is a customized dataset based on CLEVR, which contains newly generated images along with captions.

It would potentially be benefited from a fine-tuned MDETR model for detecting objects-- and to do this, I will need to obtain bounding box annotations for objects in the image, as detailed in Appendix B.1. I do have the scene files generated from Blender, which specify e.g., shape, material, and pixel coordinates of each object in the image. So it seems that the best way to get this preprocessed is for me to adopt your scripts.

It would be of great help if you could point me in the right direction here. Thanks.