GeoAi – How to handle the luxury of having too much training data

Abstract

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 training 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 training rig where we dynamically select and produce training data while training and evaluating the model - to get the best AI for building detection

Day & time

Wednesday, September 06, 5:00 – 6:00 PM

Intended audience

Anyone with an interest in AI and that would like to learn about practical use cases, and the reality behind training your own models with your own data.

• Mathilde Ørstavik

  Mathilde is the head of geospatial AI at Norkart. She has a broad experience in applied Ai on geospatial data, in particular in semantic segmentation from aerial high-resolution data.