Epipolar geometry reduces correspondence ambiguity before optimisation and improves robustness under noise.
The front-end uses the normalised epipolar residual as a gating signal before passing matches into pose estimation. This shifts work away from later optimisation stages and limits contamination from weak correspondences.