Would this method work for cases where the camera performs pure rotations in place.

[antoan]

Thank you for making this public.

I want to make an inquiry specific to my use case, where the camera rotates in place about its vertical axis, with no transnational motion - where most approaches fail due to inability to triangulate features.

I'd like to know if this approach can over come this issue?

Thanks again.

[MarkYu98]

Hello, we have not tested specifically for cases with pure rotation.

My understanding is that the hybrid estimation might not work well due to the parts based on epipolar geometry (point-based solver, triangulation, and epipolar error) do not work/degenerate with pure rotation. However, the depth-based solver and the depth-induced reprojection error should work, and I would be interesting to know if it's effective for pure rotation in your experiments :)

In addition, if you already know that the camera is performing pure rotation, I think it is also possible to develop a solver that utilizes the depth priors with that additional constraint.