AWR1642BOOST: how do I evaluate two peak grouping methods?

We don't know of any academic papers that compare the methods but here is some information on why we give these options that may help you decide:

We initially started with the detection matrix based peak grouping which seems like the simple/natural thing to do. But we noticed that this can invert the CFAR logic in that sometimes peak grouping will declare peaks that did not previously pass the CFAR test even as peak grouping is only evaluating neighborhoods of CFAR detected point but neighborhood itself is not restricted to have passed CFAR. This can happen sometimes because the two algorithms are different. We thought that some customers may not want this inversion to happen particularly if they have carefully tuned their CFAR algorithms, they may get surprised when they see output of peak grouping declaring peaks that did not pass CFAR test. The likelihood of this happening depends on the CFAR algorithm choice (CA, SO, GO) and the threshold and window size settings of the CFAR algorithm (whereas peak grouping is only trying to pick a peak in the neighborhood of the points. The inversion may happen more in the case of SO, GO which are asymmetric (compared to CA) whereas peak grouping is symmetric. On the other hand, some customers may be less emphatic/sensitive to CFAR performance and may want to make sure that if CFAR did not find peaks that peak grouping found in the neighborhood, that they may want to see those peaks. So it is hard to know who prefers what, so we decided to expose the choice in the CLI configuration and let customers decide. You can try different settings for you requirements and choose the combination that best achieves the desired performance, it is basically about practical tuning, there is no concrete theoretical guidance. Note that there are more advanced algorithms to replace peak grouping like group tracker which is available as one of the demos outside of the SDK.