Flownet2 outputs completely white flow
huberl opened this issue · comments
I'm currently trying to evaluate the flow between an object in two different images. One image is the observed camera image and the other one a rendered version of the object using a colored 3D model.
A mask of the object is added to both the observed and rendered image to yield two images with four channels.
More specifically, I'm following the approach published by:
https://github.com/liyi14/mx-DeepIM
For some reason the output of the Flownet2 is completely blank. Does somebody know the reason for this behavior?
"Zero flow" may be caused by a subtly incompatible CUDA version. Please check if #201 works for you.
It might also be worth double-checking that you are not feeding the same image twice to FlowNet2 😉
Thanks for the quick reply. I do not believe that this issue is based on the CUDA version since Flownet2 works for me on the ape model of the LINEMOOD6D dataset as shown here:
This is the output of the Flownet2 with my own 3D model. The object is centered and not in the upper right corner as the image may indicate:
Is the Flownet2 robust against 3D models that do NOT perfectly match the object in the scene by color?
Ok, that's good. FlowNet2 is definitely robust against not perfect color matches, but I never tried real vs rendered objects. Would it be possible for you to post an example image pair (or, if not, to send it to me via mail)?
(closed due to inactivity)