Ann Arbor AI, Machine Learning and Computer Vision Meetup

Building datasets for training visual AI models is laborious, error-prone, and expensive. Even worse, we lack the ability to measure how good a dataset is for a certain task. And, yet, we spend immense amounts of capital on building datasets, especially the case in the automotive sector. This talk will explore two vectors of inquiry around improving our ability to build measurably high-quality datasets for training models. First, I will explore discuss new methods in pruning large data-lakes down to size for model training guided by the ability to estimate dataset difficulty. Then, I will describe the potential for AI-human teaming when building and annotating datasets, leveraging a human in the loop to guide auto-labeling models when labeling complex scene geometry for mobility datasets.