Reproducing MaskCut & CutLER (1 round) on COCO val2017

Hello!

I have recently attempted to reproduce the numbers reported in the paper for COCO val2017, both for MaskCut and Cutler with 1 round of training, but unfortunately, the results I got were not matching what I was expecting from the reported results in the paper.
For MaskCut, I got 1.7 APmask instead of 2.2. I also ran the pseudo mask generation for all of ImageNet and trained a Cascade Mask R-CNN with your config. For 1 round of training on the pseudo masks, I got 7.3 APmask instead of the reported 8.8.

I was wondering if you could help out here to identify potential mistakes I might have made? I was closely following the instructions of the repository, also using the same torch version. I also got a message that all folders were annotated, confirmed your print statement the merge_jsons.py script.

Furthermore, it would be super helpful to have the code used for evaluate the MaskCut pseudo masks. Would it be possible for you to release this script? :)

Many thanks in advance!