Preference of displacement values to return zero
ShashankBice opened this issue · 3 comments
Hi,
I have been noticing that when I run autoRIFT with two images and there is shadow over one of them, the displacement values around the the shadow area are assigned value very close to zero. Is this a known issue and are there existing logical workarounds about it ?
I was thinking of masking off two low displacement values, but does not seem to be a clean hack to me.
Example1:
Input image:
Output displacement (m/day) (Notice the areas around the shadow in either of the input images):
Thanks,
Shashank
It looks like you have artifacts in the image or being introduced in the orthorectification. This is why you have banding in the displacement results. My guess is that when you lose features in the image due to shading these artifacts are the most prominent feature for autoRIFT to lock onto (maximum NCC). My conclusion is that your imagery may not be of high enough quality for feature tracking. You could try doing and fft decomposition, remove frequencies with abnormally large power then reconstruct the image and try again. Or you could try to improve the orthorectification or imagery quality
Hi Alex,
Thanks for your response, your high frequency removal suggestion made me question my methods a bit, and looks like using a laplacian pre filtering kernel with too small a window size was introducing those. They do not appear if I do not prefilter the images. To get rid of absolute zero displacement values in shadowed areas, I just simply apply a np.ma.masked_equal
operator for now.
I think this issue can be closed for now.
Thanks,
Shashank
Sounds good.. closing.