Active learning for segmentation based on Bayesian sample queries

Firat Özdemir, Zixuan Peng, Philipp Fürnstahl, Christine Tanner, Orcun Goksel

Abstract: Highlights•Bayesian sampling with information maximizing autoencoders for active learning.•Task agnostic representation sampling combined with task oriented uncertainty.•Iterative rapid coverage of a subset distribution with the full set in latent space.•Quantify representational power of a sample from a large high dimensional dataset.•High segmentation performance with only a small subset of the dataset been annotated.