Question: How do you sample base training dataset?

[ZhangGongjie]

Do you remove all images that contain novel category objects, or remove all instances of novel categories and then remove images with empty annotation?

[YoungXIAO13]

Hi @ZhangGongjie

These are actually two possible choices (see here).

In our initial submission, we used the second choice (which we believe is more strict). And in the post-submission experiments, we found that the first choice (which is adopted in TFA) can bring us some improvement on COCO but won't change too much on Pascal-VOC, we thus used the first choice in the current code.