tp.bandpass issues

[jab-cmd]

Hi,

I've come across two issues regarding the function bandpass of trackpy.

The first one is regarding the function vectorization.
Basically, when I use it directly on a numpy array of images without using a loop, the first images of the stack and the last one are completely black, meaning, almost every pixels is somehow clipped to 0. Only a few are at a certain random value(i.e 10 or 15 pixels out of 1024*1024). (I am working on 8bits images, each pixel ranging from 0 to 255 in value). This issue resolves itself when I use a loop in order to process the stack.

ephem=tp.bandpass(ephem, l_short, l_long, trshold)
-> First and last images are lost

vs

 for i in range(len(ephem)):
     ephem[i]=tp.bandpass(ephem[i], l_short, l_long, trshold) 

-> Every images are intact

The second situation regarding this function is about its version. I applied the current version of the function and the version from the v0.3x release on the same stack of images, with the same parameters (llong, lshort etc...) and got way better results with the version from 8 years ago. I know both versions are not going to do the exact same thing, but it looks like the new version is worse.

FIrst image is after use of tp.bandpass from latest release.

Second image is after use of tp.bandpass from release v03.x

parameters used : lshort =1,llong=5, threshold = 1
As you can see, the version from the old release looks smoother, cleaner. This results in a better tracking for us afterwards.
I would expect a newer release to be better...
What do you guys think ? Would love to have insights regarding this.

Thank you.

[nkeim]

Hi! I'm certain about your first question: bandpass is treating your stack as a 3D (volumetric) image. You need to call it individually for each frame.

I'm not sure about the second question. Have you tried using lshort=2 (or 1.5, etc.) and seeing if it's more to your liking? The simplest explanation is that at some point we changed how that length scale is measured. So the next place to investigate would be in the release notes.

I hope this helps!

[nkeim]

…and I can already predict you'll find that in v0.3 we were using FFT by default for bandpass, and I think in recent versions we use convolution kernels because they are much faster. So it could just be the usual ambiguities in converting between the scale of a kernel and a wavenumber.