Cv bench data vqa creation from ground truth coco annotations

[HariniNarasimhan]

Great work. Enjoyed your presentation at Neurips. Can you please add the scripts used to create the cv bench dataset from the ground truth annotations of coco, ade20k and Omni3d ?

this might help us to generate similar benchmark for custom dataset. Thanks in advance.

[futurev]

where can we find the 3D_VQA.json annotation?

[mashijie1028]

@futurev
I figured it out! Just download the full CV-Bench dataset from https://huggingface.co/datasets/nyu-visionx/CV-Bench and revise onmi_eval.py by filtering out 2D data like:

if line["type"] == '2D':
    continue