Stochastic Methods for Fine-Grained Image Segmentation and Uncertainty Estimation in Computer Vision
Date of Award
Fall 2020
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Electrical and Computer Engineering
First Advisor
Medeiros, Henry
Second Advisor
Tabb, Amy
Third Advisor
Odone, Francesca
Abstract
In this dissertation, we exploit concepts of probability theory, stochastic methods and machine learning to address three existing limitations of deep learning-based models for image understanding. First, although convolutional neural networks (CNN) have substantially improved the state of the art in image understanding, conventional CNNs provide segmentation masks that poorly adhere to object boundaries, a critical limitation for many potential applications. Second, training deep learning models requires large amounts of carefully selected and annotated data, but large-scale annotation of image segmentation datasets is often prohibitively expensive. And third, conventional deep learning models also lack the capability of uncertainty estimation, which compromises both decision making and model interpretability. To address these limitations, we introduce the Region Growing Refinement (RGR) algorithm, an unsupervised post-processing algorithm that exploits Monte Carlo sampling and pixel similarities to propagate high-confidence labels into regions of low-confidence classification. The probabilistic Region Growing Refinement (pRGR) provides RGR with a rigorous mathematical foundation that exploits concepts of Bayesian estimation and variance reduction techniques. Experiments demonstrate both the effectiveness of (p)RGR for the refinement of segmentation predictions, as well as its suitability for uncertainty estimation, since its variance estimates obtained in the Monte Carlo iterations are highly correlated with segmentation accuracy. We also introduce FreeLabel, an intuitive open-source web interface that exploits RGR to allow users to obtain high-quality segmentation masks with just a few freehand scribbles, in a matter of seconds. Designed to benefit the computer vision community, FreeLabel can be used for both crowdsourced or private annotation and has a modular structure that can be easily adapted for any image dataset. The practical relevance of methods developed in this dissertation are illustrated through applications on agricultural and healthcare-related domains. We have combined RGR and modern CNNs for fine segmentation of fruit flowers, motivated by the importance of automated bloom intensity estimation for optimization of fruit orchard management and, possibly, automatizing procedures such as flower thinning and pollination. We also exploited an early version of FreeLabel to annotate novel datasets for segmentation of fruit flowers, which are currently publicly available. Finally, this dissertation also describes works on fine segmentation and gaze estimation for images collected from assisted living environments, with the ultimate goal of assisting geriatricians in evaluating health status of patients in such facilities.