The scripts postprocess.py and vis.py are used to generate detections from attention maps produced from MAC-network.
You will need to update the file paths for data and prediction files according to your file structure.
The script postprocess.py will generate a file named grounding_results_corrected_{tier}_0.5_vqa_gt.json (tier='val')
I used the script compute_scores_updated_gqa.py to get the detection scores (overlap and iou in terms of P, R, F1).