Why The Hungarian Algorithm?

[SYoung97]

I was reviewing your methods of matching the inferred emitter locations with the ground truth positions. I see you are using the Greedy Hungarian Algorithm , and I wondered why you chose this algorithm, specifically instead of using the Gale-Shapley algorithm as the solution to the Stable Marriage Problem.

As far as I can tell, the Greedy Hungarian Algorithm is preferred when the most important metric to minimise is the rmse as the minimum sum of Euclidean distances of a batch of inferences will (usually) be produced. However, in instances of very dense signals, there often arise situations where minimising the sum of the Euclidian distances will not necessarily maximise the true positive rate, as defined by a region of acceptance around a ground-truth position. If the true positive index, and thereby the Jaccard index, were more important performance metrics than the rmse, should the Gale-Shapely algorithm be preferred?

I'm sure the two algorithms perform very similarly for the vast majority of experiments, but I would be very interested to know the reasoning behind your choice! Thank you very much for your time: your DECODE paper has greatly impacted my work for the better!

[ASpeiser]

We used exactly the same algorithm as employed by https://srm.epfl.ch/Challenge
You can check their publication for their reasoning.
In cases like this it is important that there is an agreed upon set of metrics so that results can be easily compared.