ComBiNet: Compact Convolutional Bayesian Neural Network for Image Segmentation

Fully convolutional U-shaped neural networks have largely been the dominant\napproach for pixel-wise image segmentation. In this work, we tackle two defects\nthat hinder their deployment in real-world applications: 1) Predictions lack\nuncertainty quantification that may be crucial to many decision-making systems;\n2) Large memory storage and computational consumption demanding extensive\nhardware resources. To address these issues and improve their practicality we\ndemonstrate a few-parameter compact Bayesian convolutional architecture, that\nachieves a marginal improvement in accuracy in comparison to related work using\nsignificantly fewer parameters and compute operations. The architecture\ncombines parameter-efficient operations such as separable convolutions,\nbilinear interpolation, multi-scale feature propagation and Bayesian inference\nfor per-pixel uncertainty quantification through Monte Carlo Dropout. The best\nperforming configurations required fewer than 2.5 million parameters on diverse\nchallenging datasets with few observations.\n