Possible Improvements
kaylode opened this issue · comments
- Try different losses (default is STCN's OhemCE)
- #25
- Choose reference based on organs's sizes or adapt active learning. Experiment other different referencing strategies.
- Experiment different propagation strategies (default is save mask every k frame into the memory, then use top m frames to predict)
- Add fusion module from MiVOS
- Add CPS (or other semi-sup models) for reference model
- User-friendly GUI application
- Uncertainty Estimation (for both training and inference)
- Balance class sampling (help improve potentially)
- Heuristic thresholding for each classes
- Boundary-awareness