Lvis 92-i

[Revist]

Thank you for very inspiring work! Our research group especially thinks that creation of new, demanding few shot dataset is a great idea. As we would like to use this dataset for evaluation we have a few questions about lvis 92-i dataset:

1. How exactly the folds are chosen i.e. which classes belong to which fold? Usually a table with all names is supplied. The methods are different for Pascal-5i where continuous segments are taken (class 1, 2, 3, 4, 5 for fold 0, then 6 .. 10 for fold 1 etc) and coco 20i where discontinuous numbers are taken (class 1, 5, 9, ... for fold 0 then 2, 6, 10, ... for fold 1 etc)
2. You eliminated classes with less than 2 images for 1 shot task. Using this logic for 5 shot one would need to eliminate classes with less than 6 images. This will change the dataset most probably. As you mentioned that you did evaluations for 5 shot scenario in the other issue what are the exact classes and folds in this scenario?
3. Would it be possible to release dataloading and evaluation code? Our group has noticed a mistake in how evaluations are done for COCO 20i and we think this ought to be avoided in case of new dataset. Namely in some prominent implementations classes are encoded as .png images with class masks having the same number as their class number (i.e. pixels of class number 1 have value 1 in the image). The problem is that if masks of two classes overlap then one mask is cut away.

Thank you very much for your response in advance!