suprising results

[ilyak93]

Hi, thank you for your work.

I try to run FADNet on thermal stereo pairs of images as proposed in here, which relates to your network as well.

I get pretty strange results, I'll give u an example:

my inputs are normalized 1d thermal images of size (512, 640):

the first is the left image, second is the right, the third is my actual disparity label which is the values got from the pixel correspondence from the left and right images (the square root of the sum of squared differences of the corresponding pixels coordinates) normalized by 256, the fourth is just same disparity but squared (I was not sure if take the square root or not, so just visualized it now), the fifth is output_net1 for the taken left and right images and the six is output_net2 for the taken images, which is also the disparity output as much as I understand.

Any reasonable explanation why this is the out put and what should be adjusted to get disparities as actual output?

P.S: both normalization (of the input to range [0,1] by dividing by 2^16 an of the disparities by 256) are made experimentally, that way the network learns well and the losses are decreasing reasonably.

P.S:
I've trained for 77 epochs, maybe it's not the final result and it will unite the double vision it has toward now.

[blackjack2015]

Thanks for the interesting application. We are trying to conduct some experiments on the thermal images.
We are also welcome any collaboration on the topic of stereo matching based on thermal information. :-)