train/val list for coco_20i folds

[silsgah]

Hi researchers, I was looking at recreating the split for the coco_20i folds, I have done that but I wish to know if there is any way of verifying this because PFENet(https://github.com/dvlab-research/PFENet) was not very clear on how the train/val folds 0,1,2,3 were obtained. Any leads would be much appreciated. The train/val lists for COCO contain 82081 and 40137 images respectively but the train/0 count was 21917 and val/0 26052 of which the total is 47969 and this number is way less than the total coco set of images of 82081. This is the total accounting of train/val by the PEFNet paper.
PFENet list folder accounting analysis
Split 0: train -21917 val 0 -26052 total - 47,942
Split 1: train -41809 val 1 -15305 total - 57,114
Split 2: train -43045 val 2 -13138 total - 56,183
Split 3: train -44614 val 3 -12640 total - 57,254.

This is what I obtained per the understanding I obtained from the paper and code base
Split 0: Train=68839, Val=10959
Split 1: Train=75780, Val=4018
Split 2: Train=75336, Val=4462
Split 3: Train=75239, Val=4559
I would appreciate very much how these coco_20i split of images were obtained.