How to generate Figure 2(b)? about igev

[gangweiX]

By function corr(fmap1, fmap2), I can get correlation [B, H, W1, W2]. along the last dim, apply softmax function to compute disparity.

[waychin-weiqin]

Thank you so much for the prompt reply! Can you clarify if to apply softmax or soft-argmax to the last dim of the correlation volume?

Thank you again :)

[gangweiX]

pseudo code：

CV : [B, H, W1, W2]

prob = F.softmax(CV, dim=-1)
disp = torch.arange(0, W2, dtype=prob.dtype, device=prob.device)
disp = disp .view(1, 1, 1, W2)
disp = torch.sum(prob * disp , dim=-1)
coords = torch.arange(W2).float().to(disp.device).reshape(1,1,W2).repeat(B, H, 1)
final_disp = coords - disp

I wrote it by hand, without debugging, maybe it has some mistakes, but the process is like this