Question about ill-posed regions.

[JarvisLee0423]

Thanks for your excellent work!
I have some questions about the ill-posed regions.
First, I want to make sure I correctly understand the ill-posed regions. Are the ill-posed regions in stereo matching mainly talking about the occlusion (including non-overlap regions) and texture less regions?
Second, why you said in your paper that the ill-posed regions are mainly depend on long-range and corse-grained matching similarities? If the ill-posed regions are the occlusion and texture less regions, in my view, the occlusion regions can not get the correct information from the matching similarities no matter it was computed by long-range similarity or short-range similarity, cause the occlusion regions can only be seen in left image, which means these regions can not find matched pixels in right image. As for the texture less regions, all the pixels are quite similar, therefore, in my view it is hard to find its own matching point in right image no matter in long-range similarities or short-range similarities. Therefore, I am a little bit confused for your selective geometric feature fusion module, why it can be used to tackle the ill-posed regions, which just use left feature as a gate to fusion the multi-range similarities?

The other main reason that I ask these questions is that I actually revised your evaluation codes for sceneflow dataset to separately test occlusion and non-occlusion regions performances, where the occlusion mask is computed by thresholding the warping error between left and right images via gt-disparity. I actually noticed that the epe and d3 in occlusion regions are quite lower than other iteration-based methods. For the non-occlusion regions the d3 is lower than other methods, but the epe is similar with other methods. It means your models do tackle the occlusion regions, but I still confused why it can work like this theoretically.

Hope for your kind and early reply!