Seed funding: Feasibility study of MacDon datasets for machine learning development

This project aims to determine if MacDon’s existing data can be used to develop machine learning models for image segmentation. Image segmentation means identifying different objects in an image by assigning each pixel to a specific category. MacDon has lots of unlabelled video and image data collected from their farming equipment. Traditionally, large amounts of labelled data are needed to create effective machine learning models. MacDon tried to create synthetic (artificial) data similar to their real field data, but the models trained with this synthetic data did not perform well on actual field data. In this project, we’ll analyze both MacDon’s real and synthetic data, as well as their current models, to find out what improvements are needed. We will look at the quality, relevance, size, variability, and noise in the datasets. We’ll also investigate the consistency of the labels in the data. For the models, we will examine why they did not meet performance expectations and suggest improvements. The goal is to provide a detailed report with recommendations on how MacDon can improve their data and models for better image segmentation, laying the groundwork for future collaboration and development.

Faculty Supervisor:

Christopher Henry;Shaowei Wang

Student:

Partner:

MacDon Industries Ltd.

Discipline:

Computer science

Sector:

Manufacturing

University:

University of Manitoba

Program: