Active Learning for Segmentation Based on Bayesian Sample Querying

A representativeness metric proposed for the acquisition function of an active learning framework.

We release code and demo scripts for our work on querying new samples to maximize representativeness within a dataset. Within the scope of semantic segmentation, our proposed approach vastly reduces the amount of necessary annotated samples to achieve performance on par with having the full dataset annotated.

Code: github.com/firatozdemir/al-bsq