Probabilistic two-stage detection

March 2021

tl;dr: How to build fast and accurate two stage detector on anchor-free one-stage detectors.

Overall impression

It seems to me that the “probabilistic” looks fancy, but actually are not that appealing for practical use. The main achievement of this paper is on how to extend SOTA one-stage detectors for two stage and make them fast.

First stage of most two stage detectors is to maximize recall.

Current SOTA one stage models rely on heavier separate cls an reg branches than two-stage models. They are no longer faster than two stage methods if the cls (vocabulary) is large.

Previous two stage methods are slow due to the large number of proposals in the first stage. And two-stage methods

Key ideas

Technical details