Closed
Description
Currently, proofreading is primarily used to correct track identities to fix swaps and etc.
To adjust one or more landmarks, you have to double click on the prediction to create a user instance from it with nodes that are movable.
The problem is that both the original predicted instance and the new (corrected) user instance are both still stored in the labels. This could result in the original predicted instance being used instead of the new user instance when exporting to an analysis HDF5 file.
Thanks @olinesn for raising this issue!
Solutions:
- We should explicitly check for linked predicted <-> user instances by inspecting the
Instance.from_predictedattribute. - We could just write out all the predicted instances to the appropriate tracks and then the user instances, overwriting any predicted instances in the corresponding track channels of the exported tracks array.
The first option is more general and would work even when the Instance.track attribute isn't set (e.g., for single instance videos).