Abstract
Recent studies have shown that ensemble approaches could not only improveaccuracy and but also estimate model uncertainty in deep learning. However, itrequires a large number of parameters according to the increase of ensemblemodels for better prediction and uncertainty estimation. To address this issue,a generic and efficient segmentation framework to construct ensemblesegmentation models is devised in this paper. In the proposed method, ensemblemodels can be efficiently generated by using the stochastic layer selectionmethod. The ensemble models are trained to estimate uncertainty throughBayesian approximation. Moreover, to overcome its limitation from uncertaininstances, we devise a new pixel-wise uncertainty loss, which improves thepredictive performance. To evaluate our method, comprehensive and comparativeexperiments have been conducted on two datasets. Experimental results show thatthe proposed method could provide useful uncertainty information by Bayesianapproximation with the efficient ensemble model generation and improve thepredictive performance.