Meta learning approaches for few-shot semantic segmentation with sparse labels

• Autor(a) principal: Pedro Henrique Targino Gama
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Institucional da UFMG
• Texto Completo: http://hdl.handle.net/1843/58285; https://orcid.org/0000-0002-9802-593X
• topic: Computing; Computer Vision; Machine Learning; Deep Learning; Meta Learning; Computação – Teses; Visão por computador – Teses; Aprendizagem de máquina – Teses; Aprendizado profundo – Teses; Meta-aprendizado – Teses

Semantic Segmentation is a classic task in Computer Vision that has multiple applications in many areas, from organ segmentation for clinical image studies, or counting objects in production lines, to estimating deforestation areas sizes. However, the type of data labeling required for actual methods to solve this problem is laborious to produce, since one has to determine the label for all pixels of an image. This usually increases the cost (human and/or monetary) to produce new datasets. Two possible ways to reduce this cost are: 1) reducing the number of labeled samples; 2) using simpler/sparse types of annotation. Despite that, current and usual deep learning based methods for segmentation tend to perform poorly when using one, or two, of these solutions. In this work, we propose two meta learning methods to the few-shot semantic segmentation task with sparse annotations. These two approaches are based on two existing methods for classification: Model-Agnostic Meta-Learning (MAML) and Prototypical Networks. Our methods were tested in different scenarios in the medical and remote sensing areas, which usually have limited data access, and obtained competitive results in different tasks.