What is the loss of 2D Affine Motion Estimation? Answer: diff between predicted and truth motion about deeperinversecompositionalalgorithm HOT 3 CLOSED

Answer: diff between predicted and truth motion, also, it is 2d Affine Motion, not 3D motion,
3D motion must be together with depth or depth prediction

It seems that all optimization-based motion estimation methods are supervised learning.
Because it optimizes the pose directly and iteratively, eg. IC-STN, DEEPV2D.
Can we set the last iterative error delta_efslu to be zero as the unsupervised learning metric?
Also, an additional photometric loss or feature metric loss based on the last iterative residual r could also be added.

from deeperinversecompositionalalgorithm.

Awesome repo and paper!
It is not a research, it is a lecture!

from deeperinversecompositionalalgorithm.

I did not release the 2D affine transform, which should be straight-forward to implement according to the supplementary if there is a need. I tried to make this repo as concise as possible. I am glad it helps you.

I would not say direct supervision is necessary, but empirically I would say it ensures there is no ambiguity in the optimization. If you have tried to train yourself, you may find it is not that easy to train the unrolled optimization. How to make it faster to train, have better convergence, and use self-supervision are all interesting open research questions. I would like to know more if you have some findings in these.

from deeperinversecompositionalalgorithm.