About stereo-matching backbones

[ahtso]

Thanks for your contribution, it helps me a lot.

I have a question about stereo matching network backbone.
Since your experiments, you use PSMNet, GwcNet, GANet and PCWNet on github.
I was curious about why not using ACVNet or Raft-stereo (Recurrent based) related?
Is there any reason?
Or something I missed in paper.
If I wanna use the ADL module, the network must be cost volume based right?

I want to take perfomance and time into account, any suggestion of the backbone?
Thank you for your answer.

--- New Question ---

I tried to use PCWNet as my backbone, and used ADL module.
After I reviewed the PCWNet backbone,
there is a refinenet, and it required that makes the disparity shape like ( Batch, 1, H, W )
but when use ADL module, it required that disparity shape like ( Batch, disp, H, W) to evaluate the cross entropy loss.
I have no idea to output the backbone output like ( Batch, 1, H, W ) or ( Batch, disp, H, W )

Hope you reply soon, Thank You A lot.