At Norkart we aim to develop the best AI models for automatic mapping of objects from aerial imagery. With a wealth of already labeled objects, such as buildings, we find ourselves in a somewhat unique position - we have too much training data! I know what you’re thinking! There’s no such thing as too much data. However, a large amount of irrelevant data can impede the development of a well-balanced training dataset. For example, when training a building-detection model, we need to be selective in the examples we use, focusing on a diverse range of building types rather than non-relevant data such as oceans, forests, and roads. So how can we ensure an ideal selection of training data in order to get a model that is robust enough to analyze any part of Norway and recognize any sort of building? Join us as we present our approach to dynamically selecting and producing training data while training and evaluating the model - to get the best AI for building detection. Check out our new channel: NDC Clips: @ndcclips Check out more of our featured speakers and talks at https://ndcconferences.com/ https://ndcoslo.com/