proxy disparity generation

[xjturobocon]

Hey, I tried to use the method "Unsupervised-Confidence-Measure" to generate proxy disparities. However, I found it can only get left disparity. How did you generate the right one?

Another question is that sampling probability, also the argument "input_points" in main.py. I see it is different between train and test in your paper. Specially, 1/1000 in train stage, 1/20 in test. But you didn't set it in the readme. I want to confirm it.

hope for your reply.

[fabiotosi92]

Hi, in order to generate the right disparity map using "Unsupervised-Confidence-Measure" given an input stereo pair, you can flip horizontally the left and right images and then switch them before passing to the method. To put it simply, the horizontally flipped left image become the new right image and the horizontally flipped right image become the left one. As final step, you just have to flip again the disparity outcome from the method. By doing so, you will have a disparity map aligned with the right view of the initial stereo pair.

For what concern the "input points", we confirm that we use 1/1000 during training and 1/20 during testing in our experiments. You can find such implementation detail in Section 4.2 of the main paper.

[ksh11023]

@xjturobocon

Hello,
Have you reproduced this repo?