The winner-take-all selection

[DIVE128]

Hellow, thanks for your great work and contribution. The unification applies winner-take-all selection as the basis to generate the depth map for the next finner stage. However, torch.max() is a non-differentiable operation and the parameters in 3D-CNN would be optimized except the parameters w.r.t the selected depth hypothesis. How do the gradients backward in this case?

[prstrive]

You are right! So we optimize the model through directly constraining the cost volume, but not the depth. You can refer to our paper or here for more details.

[DIVE128]

Thanks a lot.