The epipolar_loss not converge

[EAST-J]

Hi, I tried using the epipolar_loss on my data to refine a noisy pose, but it doesn't seem to be converging. Could you provide some guidance? Could this be due to a poor initial pose or inaccurate correspondences?