Object detection model training

Abstract

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a model to detect objects. One of the methods includes maintaining first training data that represents at least a plurality of first images of invertebrate fossils; determining, using at least one or more geographic properties of a geographic region, one or more transformations from a group comprising: a pixel group translation, a pixel group two-dimensional rotation, a changed background, or an object obfuscation; creating, using the one or more transformations and the first training data, second training data that represents at least a plurality of second images of invertebrate fossils; training, using the second training data, a machine learning model to detect invertebrate fossils in images of the geographic region; and providing, to a system, the trained machine learning model to enable the system to detect invertebrate fossils in images of the geographic region.