ScaViSLAM
ScaViSLAM copied to clipboard
Validate/improve sub-pixel accuracy
At the moment there is a method for sub-pixel accuracy at the moment, which - however - is deactivated. It is a simple direct method (Lukas Kanade tracking) using 1 iteration of Gauss Newton.
There are a few things to do:
- Test whether everything works correct.
- Try out ESM and see whether it pays off.
- We should not include accept updates < 1 pixel.
- Shall we test whether the update actually reduces the error (either using LM or line search)?