Taking derivatives of a noisy time series of angles is always a huge pain (mostly because unwrapping noisy data is very hard). It would be nice if we could provide some examples (maybe helper functions?) of how to do this, which methods work best, etc. Furthermore, it could be that some methods (like the loss based methods) can be adjusted to do this in a much more elegant way by implementing an angular difference in the loss function.
Discrete action items:
- Band-aid: provide a helper function to unwrap and then differentiate angular time series
- Try modifying a loss based method to do angular losses, and if that works, extend to other loss based methods
- Decide on a software architecture that would be clean. For example, would angular helper functions and modified loss methods live under their own "angulardiff" subpackage, or do we modify methods so that they have an "angular" flag?
Taking derivatives of a noisy time series of angles is always a huge pain (mostly because unwrapping noisy data is very hard). It would be nice if we could provide some examples (maybe helper functions?) of how to do this, which methods work best, etc. Furthermore, it could be that some methods (like the loss based methods) can be adjusted to do this in a much more elegant way by implementing an angular difference in the loss function.
Discrete action items: