Consistency on the Vertices, Edges and Mask: A Semi-Supervised Learning Approach for Image Segmentation

Abstract

In this thesis, we present a novel method for performing image segmentation in a semi-supervised approach, which we consider to be particularly relevant because of the substantial cost of obtaining pixel-wise annotations required to train supervised. This method, that we will call Consistency on the Vertices, Edges and Mask, is one of the first methods that can be used for training deep neural networks to perform image segmentation in a semi-supervised setting where only a small portion of training data is labeled. In our setting, we train a network to predict masks, edges and vertices for a given input image, and then we penalize the network for not being consistent with its predictions obtaining the theoretical edges an vertices from the predicted mask using a derivable version of the Canny edge detector. We also present results on the Cityscapes Dataset where we obtain outstanding results achieving about 92% of the fully supervised performance labeling only 10% of the data and 98% labeling 25%.

This thesis also contains a brief introduction on the field of deep learning and semi-supervised learning, relevant previous work that has been published in the last year which inspired