What exactly is supposed to happen with iscrowd bounding boxes during object detection training?

Question

I have two questions regarding the iscrowd tag in object detection:

1. Let's say you have a crowd of objects. You are able to label some of them individually. When you have such a situation should you:

   a) label everything you can AND THEN additionally create a big iscrowd bounding box covering the entire crowd?

   b) create a single iscrowd bounding box and not even bother labeling objects even if it's possible for some of them?

   c) label objects that can be labeled individually and cover only the rest with one singular (smaller than in (a)) iscrowd tag?

2. How to use iscrowd bboxes in training?

   a) Treat them like normal bboxes?

   b) Discard them completely?

   c) Discard them only if you have some other bounding boxes from the same class with non-zero IoU with the iscrowd bbox?