​​Let's talk a bit about object detectors.
But we want to have a single bounding box per object (not hundreds of them), right?
You probably know that in most detection pipelines there is a step called Non-Maximum suppression (NMS) which is responsible for this.
Its purpose is to get a lot of tentative detections proposed by the networks and drop all spurious and highly overlapping ones, retaining only a single the most confident bounding box per object.

The de facto standard approach for NMS is Greedy NMS. At each step, we select the most confident box and drop all others which have IoU greater than some fixed threshold (often 0.5). We repeat this process until no proposals are left. But this approach is very taxing and requires manual tweaking (which I personally hate). For example, if you set a threshold too high you may lose Recall, since some very close objects would be considered as duplicate detections and would be dropped. On the other hand, a lower threshold may leave you with too many spurious detections.

Therefore researchers from Max Planck Institute for Informatics in Saarbrücken proposed a method for end-to-end learnable NMS. Now we train a neural network to do NMS instead of a greedy algorithm. Details are in the paper: 📰 "Learning non-maximum suppression"

I briefly summarize the proposed algorithm, but for more detail refer to the paper.
If the object is already assigned to one detection, all other detections with the high overlap (neighboring detections) should be notified about it and should decrease their scores. To do this the paper proposed to compute pairwise features between overlapping proposals (these features are handcrafted and include IoU, normalized distance in X and Y directions, a difference of width and height, aspect ratio difference, detection scores, etc.). These pairwise features are concatenated with original detection features produced by the CNN backbone and are passed through a series of residual blocks with FC layers (see Figure). Next, to assign only a single detection per object, authors run a Hungarian matching algorithm between GT and detections and enforce all non-matched detections to decrease their scores. After that, the proposed network, called GNet, can produce only a few boxes with very string scores, and all other boxes are assigned very teeny ones (see examples below).