search

princeton-vl / Coupled-Iterative-Refinement

MIT License
104 stars 20 forks source link

Implementation details about BD-PnP. #1

Closed JiChun-Wang closed 2 years ago

JiChun-Wang commented 2 years ago

Thanks for your provided code. I don't understand why you used Ji when computing the Jacobian matrix about G0 as follows: https://github.com/princeton-vl/Coupled-Iterative-Refinement/blob/fa7eed5cc1728705bf373c5b65283284b5d6f859/utils/geom/projective_ops_rgbd.py#L91 I'm new to Lie group theory, and as far as my knowledge is concerned, Jj seems to be the Jacobian needed. Hope for your help.

lahavlipson commented 2 years ago

This is true when Jj corresponds to the pose of the object in the image, rather than a rendered view. However, we also compute flow in the opposite direction, in which case Ji corresponds to the pose of the object in the image. The Js corresponding to the rendered views all get discarded here just before solving for the update.

lahavlipson commented 2 years ago

I am closing this issue. Feel free to re-open.