is there some thing with your code ”compute_epipolar_error“?

[lvzhengyi0204]

Hello：
I am looking at the source code, but found some problems。

https://github.com/magicleap/SuperGluePretrainedNetwork/blob/master/models/utils.py#L355
In L367 ：E = t_skew @ T_0to1[:3, :3] but the essential matrix is wrong. the correct code is
E = T_0to1[:3, :3] @ t_skew

and In L369 L370, the code is also wrong
Ep0 = kpts0 @ E.T # N x 3
p1Ep0 = np.sum(kpts1 * Ep0, -1) # N

the correct code is
Ep0 = E@kpts0.T
p1Ep0 = np.sum(kpts1.T@Ep0,-1)

I don't know why you wrote that. Can you explain it?

[sarlinpe]

Thank you for your interest in our work. These are not issues :)

• Let's assume that the epipolar constraint is p1 @ E @ p0 = 0
• In our code, R and t are such that they transform from 0 to 1: X1 = R @ X0 + t (so t is in the coordinate frame of 1)
• From Section 9.6 in Hartley & Zisserman: E = skew(t) @ R

I know that Wikipedia and Longuet-Higgins write E = R @ skew(t), but they define t such that X1 = R @ (X0 - t).

The code computing p1Ep0 is correct, please check before asking.

[lvzhengyi0204]

thank you very much！