Deep learning methods with less supervision

Abstract

To tackle the immense burden of acquiring accurate, pixel-level annotations for semantic segmentation tasks, we propose a weakly-supervised deep learning framework. We incorporate state-of-the-art foundational models to propagate pseudo-labels. Then, explore the viability of training a fully convolutional network based on our pseudo-labels. In addition, we experiment and evaluate the results of different loss functions and attempt the refinement of masks using conditional random fields.