Weaklier Supervised: Semi-automatic Scribble Generation Applied to Semantic Segmentation

Abstract

With many applications regarding semantic segmentation arising, along with the advent of the Deep Semantic Segmentation Networks, the need for large labeled datasets has also largely increased. But labeling thousands of images can be very expensive and time-consuming. Approaches such as weak and semi supervision try do deal with this problem, but the first cannot deal with large datasets and the latter is hard to deal with semantic segmentation. Therefore, in this work we propose a combination of both to create a novel pipeline of weak supervision, with focus in satellite imagery, capable of dealing with large datasets. We propose a pipeline to automatically generate scribbles in images, requiring that the user only label 10% of the images in a given dataset, while a classifier deal with the remaining images. Along with that, we also propose a simple semantic segmentation pipeline, that uses only images with scribbles to train a network. Results show that performance is lower, but similar to a fully supervised pipeline.