objective annotate datasets faster
for that:
- develop annotation tools (IIS)
- better train models (active learning, etc)
problem: generalization to new data (if we had a model for unlabeled data, we could use it to annotate)
So far:
- lit review: click-based deep learning methods (gto99, ritm)
- iislib: allows for experimentation (to be shown next)
- 2 mvasat projects
- igarss
In progress:
- correct igarss
- IPOL 1: clicking procedure comparison
- IPOL 2: mvasat1, generalization to aerial images
Future:
- mvasat2, continual adaptation to new image sequences
- attempt to mathematically formalize the problem
- optimal clicking procedure + active learning
- use robust models (e.g. SSL trained MAE). Ideas:
- train with more similar data (mvasat1)
- train more general models (SSL)
- efficient pretraining (use classic segmentation as labels, then use learned segmentations as labels)
- mix eff pretraining with clicking, using different segmentation methods