Optimization for Image Segmentation

Abstract

Image segmentation, i.e., assigning each pixel a discrete label, is an essential task in computer vision with lots of applications. Major techniques for segmentation include for example Markov Random Field (MRF), Kernel Clustering (KC), and nowadays popular Convolutional Neural Networks (CNN). In this work, we focus on optimization for image segmentation. Techniques like MRF, KC, and CNN optimize MRF energies, KC criteria, or CNN losses respectively, and their corresponding optimization is very different. We are interested in the synergy and the complementary benefits of MRF, KC, and CNN for interactive segmentation and semantic segmentation. Our first contribution is pseudo-bound optimization for binary MRF energies that are high-order or non-submodular. Secondly, we propose Kernel Cut, a novel formulation for segmentation, which combines MRF regularization with Kernel Clustering. We show why to combine KC with MRF and how to optimize the joint objective. In the third part, we discuss how deep CNN segmentation can benefit from non-deep (i.e., shallow) methods like MRF and KC. In particular, we propose regularized losses for weakly-supervised CNN segmentation, in which we can integrate MRF energy or KC criteria as part of the losses. Minimization of regularized losses is a principled approach to semi-supervised learning, in general. Our regularized loss method is very simple and allows different kinds of regularization losses for CNN segmentation. We also study the optimization of regularized losses beyond gradient descent. Our regularized losses approach achieves state-of-the-art accuracy in semantic segmentation with near full supervision quality.