Intermediate ground truth mask for tracked objects

[ShijieVVu]

Is there a way to provide intermediate ground truth mask for objects already in the memory to correct their drifting?

[hkchengrex]

For the GUI, it should be obviously doable.
For batch processing (eva.py), I haven't tried this but I think if you

1. create your new test dataset (in inference/data/test_datasets.py), which feeds use_all_mask=True to VideoReader and
2. pass exhaustive=True to mapper.convert_mask() in eval.py

, it should work.

If the drifting is severe, you might also want to flush the memory bank when a new frame is loaded. Let me know if it works.