Generating Bounding Box Supervision for Semantic Segmentation with Deep Learning

Simone Bonechi, Paolo Andreini, Monica Bianchini, Franco Scarselli

Abstract: Most of the leading Convolutional Neural Network (CNN) models for semantic segmentation exploit a large number of pixel–level annotations. Such a human based labeling requires a considerable effort that complicates the creation of large–scale datasets. In this paper, we propose a deep learning approach that uses bounding box annotations to train a semantic segmentation network. Indeed, the bounding box supervision, even though less accurate, is a valuable alternative, effective in reducing the dataset collection costs. The proposed method is based on a two stage training procedure: first, a deep neural network is trained to distinguish the relevant object from the background inside a given bounding box; then, the output of the network is used to provide a weak supervision for a multi–class segmentation CNN. The performances of our approach have been assessed on the Pascal–VOC 2012 segmentation dataset, obtaining competitive results compared to a fully supervised setting.