Weakly-Supervised Learning of Semantic Segmentation using Deep Convolutional Nerual Network

Abstract

Semantic segmentation is one of the fundamental computer vision problem that aims to assign dense semantic labels to every pixels in the image. Although recent approaches based on deep learning have achieved substantial improvement over traditional methods, training a deep neural network requires a large number of fine-quality segmentation annotations, which is a bottleneck to scale up the task to large-scale problem involving many semantic categories. To address this issue, I will present semantic segmentation algorithms based on deep learning, which are trainable under various degrees of supervision. Specifically, I will first describe a model for semi-supervised semantic segmentation, which employs combined training sets of a small number of pixel-wise segmentations and a large number of image-level class labels. Then I’ll extend the model to weakly-supervised semantic segmentation, which requires only image-level class labels for training. Since image-level labels are insufficient to provide strong supervision for semantic segmentation, I propose two independent approaches that exploit additional cues for segmentation–segmentation annotations obtained from different semantic categories and videos collected from web–which are easy to collect without direct human intervention. Extensive experiment results on PASCAL VOC benchmark dataset demonstrated that the proposed approaches substantially improved the semantic segmentation performance over comparable algorithms using the same degree of supervision.