In ~2006 the FAST corner detector was introduced. One of it characteristics is that you have to train a decision tree from some test image in order to learn the types of corners present and relevant in the test image. One of the consequences is that the images in which you will be applying the detector have to be somewhat similar in relation to the pose (perspective), global illumination, etc. to the test image used in the training phase.
Feature descriptors such as BRIEF and ORB make use of FAST or a modified version of FAST in their reference papers.
Then, in ~2010 the AGAST corner detection was proposed to address this issue. According to the authors, you should not have to retrain the detector on a test image (but you may if you really want to or think this is really necessary).
Then feature descriptors such as BRISK and FREAK showed up, using a modified version of AGAST in their reference papers.
My question is: Since AGAST is an evolution of FAST aimed at solving some of its issues, should someone simply use AGAST from now on when building a detection application? (I am not speaking of doing pure research on the topic of detectors)
Is there some paper proving or at least showing convincingly that you should now use AGAST? I know I could simply test them both every time I build a detection application, but honestly, if there is a relatively good point I should use AGAST, it would be time saving.