The Context
From all of the problems I have worked with in computer vision, the most challenging one is the object detection. This is not because the problem itself is complex to understand or bad formulated. But because we need to inject some strong priors about how we understand the world. Those priors are the anchors (which are priors about object shapes, aspect ratio...).
This prior information, although very simple to understand, it is very hard to inject on the training logic. Hence making the computation of the ground truth very messy and prone to errors. It is even harder when different object detection backbones propose different ground truth computation methods.
The Question
From mid-2019 till now there is a growing trend on research about one-stage object detectors that do not rely on anchors: hence dropping the costly NMS postprocessing and in some cases even the IoU computation. I would like to do a proof of concept with some of them so here is my question:
What are some good object detectors that do not use anchors? Or said in other words, what are the go-to object detectors for this new research trend?