Open
Description
Supporting IterDet would sound like a good idea, which works for both 1- and 2-stage detection models.
Motivation
It seems like a powerful post-processing/joint-training step with crowded scenes and densely packed object detection.
Related resources
Original article: https://arxiv.org/pdf/2005.05708.pdf
GitHub code w. existing adaptations from mmdet: https://github.com/saic-vul/iterdet
Additional context
What would be the best stage to introduce this?