Efficient Ensemble Model Generation for Uncertainty Estimation with Bayesian Approximation in Segmentation

Recent studies have shown that ensemble approaches could not only improve\naccuracy and but also estimate model uncertainty in deep learning. However, it\nrequires a large number of parameters according to the increase of ensemble\nmodels for better prediction and uncertainty estimation. To address this issue,\na generic and efficient segmentation framework to construct ensemble\nsegmentation models is devised in this paper. In the proposed method, ensemble\nmodels can be efficiently generated by using the stochastic layer selection\nmethod. The ensemble models are trained to estimate uncertainty through\nBayesian approximation. Moreover, to overcome its limitation from uncertain\ninstances, we devise a new pixel-wise uncertainty loss, which improves the\npredictive performance. To evaluate our method, comprehensive and comparative\nexperiments have been conducted on two datasets. Experimental results show that\nthe proposed method could provide useful uncertainty information by Bayesian\napproximation with the efficient ensemble model generation and improve the\npredictive performance.\n