Currently, the MRAE measurement handles the zero values in ground truth images by adding the "np.finfo(float).eps".
However, this value is too small (i.e., about 10e-16) to handle huge loss/error when a certain predicted pixel is not exactly the ground truth zero, using the relative loss (e.g., MRAE).
Posted by: ChangC @ July 24, 2018, 7:54 a.m.This bug may result in the failure of training with MRAE loss and significantly affect the evaluation of testing using MRAE metric.
It should be better to replace the "np.finfo(float).eps" (~2.2e-16) to a larger value such as 1/65536.
Posted by: ChangC @ July 24, 2018, 8:04 a.m.Hi,
Thank you for noticing this. You are correct. I did change this value for track2 but apparently, I missed it for track1. This applies to all the participants so it shouldn't affect the ranking. However, I will rectify this as you suggested.
Thanks,
Mehrdad
Hi,
Thank you for noticing this. You are correct. I did change this value for track2 but apparently, I missed it for track1. This applies to all the participants so it shouldn't affect the ranking. However, I will rectify this as you suggested.
Thanks,
Mehrdad
Got it. Thanks for you reply.
Posted by: ChangC @ July 25, 2018, 11:43 a.m.