Leverage gt poses

[rzhevcherkasy]

Hi, thanks for your great work. I'm wondering if there is any way to leverage the gt poses and only refine the predicted depth?

[MarkYu98]

Hi.

If by "refine the predicted depth" you mean to find the optimal relative scale $$\alpha$$ and the two shift values $$\beta_1, \beta_2$$ in our formulation between the two predicted depth maps, I think there's mainly two options you can do:

• solve for optimal $$(\alpha, \beta_1, \beta_2)$$ directly using least-square methods with all point correspondences and the depth maps (by fitting to the depth of triangulated 3D points, or minimizing reprojection errors)
• or find it through RANSAC estimation to be more robust from noise and outliers. In this case the minimal solver would require a single pair of (p1, p2, d1, d2) correspondence and you can solve for candidate $$(\alpha, \beta_1, \beta_2)$$ via simple triangulation. Because you are fixing the camera pose, the epipolar error is also fixed so you may use only the depth-induced projection error as your scoring.