Using EPro-PnP to solve relative pose estimation
zhirui-gao opened this issue · comments
Hi, I'm sorry to disturb you!
A problem has been bothering me for a long time. If I have 2d and a 3D correspondence, can I use Epropnp to solve the image pose ? I found that the calculated pose changes very small compared to the initial value. It is because there are too few iteration steps in the solver?
EProPnP6DoF(
mc_samples=512,
num_iter=4,
solver=LMSolver(
dof=6,
num_iter=5,
init_solver=RSLMSolver(
dof=6,
num_points=16,
num_proposals=4,
num_iter=3))).cuda()
Bestes wishes.
The PnP solver returns the relative pose (transformation) from the object to the camera, so the inverse transformation is just the camera pose relative to the object (or a world coordinate system).
I recommend setting more than 10 iterations in the LMSolver, if you are not confident
Thanks a lot!