Bounding boxes for weakly supervised segmentation: Global constraints\n get close to full supervision

Abstract

We propose a novel weakly supervised learning segmentation based on several\nglobal constraints derived from box annotations. Particularly, we leverage a\nclassical tightness prior to a deep learning setting via imposing a set of\nconstraints on the network outputs. Such a powerful topological prior prevents\nsolutions from excessive shrinking by enforcing any horizontal or vertical line\nwithin the bounding box to contain, at least, one pixel of the foreground\nregion. Furthermore, we integrate our deep tightness prior with a global\nbackground emptiness constraint, guiding training with information outside the\nbounding box. We demonstrate experimentally that such a global constraint is\nmuch more powerful than standard cross-entropy for the background class. Our\noptimization problem is challenging as it takes the form of a large set of\ninequality constraints on the outputs of deep networks. We solve it with\nsequence of unconstrained losses based on a recent powerful extension of the\nlog-barrier method, which is well-known in the context of interior-point\nmethods. This accommodates standard stochastic gradient descent (SGD) for\ntraining deep networks, while avoiding computationally expensive and unstable\nLagrangian dual steps and projections. Extensive experiments over two different\npublic data sets and applications (prostate and brain lesions) demonstrate that\nthe synergy between our global tightness and emptiness priors yield very\ncompetitive performances, approaching full supervision and outperforming\nsignificantly DeepCut. Furthermore, our approach removes the need for\ncomputationally expensive proposal generation. Our code is shared anonymously.\n