Base on this discussion, https://competitions.codalab.org/forums/24715/4935/, the ramaps
are normalized, is there any reason we need to normalize it? Based on "It normalizes all the magnitude of complex numbers between 0 and 1.",
then for any out of distribution data, you may always get false positive(like the last example shown in original paper). Thank you!
The reason for the normalization is because we need to feed data into CNNs. This preprocessing is good for network training. Before normalization, the values of the radar data can reach a very large value (10^4). Since the normalization is only done in magnitude, there will be just very little information lose. Besides, the normalization is not related to the "false positives".Posted by: ywang26 @ Feb. 15, 2021, 10:57 a.m.
what normalization is used? Is it a min-max scaler?Posted by: mrhuangchuan @ Feb. 23, 2021, 12:53 a.m.
There is only an upper bound during the scaling. You can assume min to be zero. This means, during the normalization, each complex number (both real and imagery) is divided by a scaler.Posted by: ywang26 @ Feb. 23, 2021, 12:34 p.m.
Could you please provide us the scaling values you have used to normalize both the real and imaginary parts of the radar signal?
It could be really usefull to recover the original signal. Moreover, if you have used different scaling values for the real / imaginary parts, it is not possible to compute the module of the signal.
Thanks for your answers.
@Authur, interesting idea. Are you going to extract speed from it?Posted by: mrhuangchuan @ March 5, 2021, 9:27 a.m.