Possibly typo.

[kyuhyoung]

Thanks for the great codes.
I am not sure but the following might be typo.
0 is for x
1 is for y.
x is for W
y is for H.
So the H of the 1st line and W of the 2nd line should be swapped in this normalization ?

EDGS/source/corr_init.py

Line 392 in eb21ad5

kptsA_np[:, 0] = kptsA_np[:, 0] / H * 2.0 - 1.0

kptsA_np[:, 0] = kptsA_np[:, 0] / H * 2.0 - 1.0
kptsA_np[:, 1] = kptsA_np[:, 1] / W * 2.0 - 1.0
kptsB_np[:, 0] = kptsB_np[:, 0] / W_B * 2.0 - 1.0
kptsB_np[:, 1] = kptsB_np[:, 1] / H_B * 2.0 - 1.0

[dimakot55]

Thanks for pointing this out — it is indeed a confusing part of the code.

I double-checked the coordinate conventions. In the current code path, kptsA_np is not in the usual (x, y) convention at that point. Because of how the source RoMa grid is constructed earlier, the first source coordinate corresponds to the row / height direction and the second to the column / width direction. This is why the normalization for kptsA_np uses H for the first coordinate and W for the second, followed by kptsA_np[:, [1, 0]] before triangulation.

So I think swapping only those two denominators would actually make the current code path incorrect. That said, the code is definitely confusing and should be rewritten to make the convention explicit, e.g. by naming the variables as row/col and then constructing final (x, y) normalized coordinates before triangulation.

Thanks again for catching this — we will clean this part up to avoid ambiguity.