DECODE on SMLM challenge data

[NaAMue]

I tried out DECODE on the data from the SMLM challenge (https://srm.epfl.ch/srm/dataset/challenge-3D-simulation/MT0.N1.LD/index.html). When comparing the ground truth to the localizations retrieved by DECODE, I noticed, that I need to transform the coordinates in order for them to overlay. The transformation is (y+25nm, x+125nm, -z). I read in posts from the discussion forum, that the swap of x and y and the minus in z is due to choices in the coordinate representation, but what about the shift? As the pixel size is a multiple of 8 (network input requirements), it should not be a result of the cropping. Is it that “thing” mentioned in issue #81? Is the shift always the same, that is y+25nm and x+125nm or y+0.25 pixel and x+1.25 pixel or is it different for every newly trained network?
Is there the possibility to group positions in DECODE?

[Haydnspass]

@NaAMue That is correct, the coordinates need to be converted. Any other bias is due to the arbitrary definition of the PSF centre which I think the challenge hosts correct for, but of course you can correct for it yourself on the training data.