Methods for generating a dataset of corresponding images for machine vision learning

Abstract

Machine learning vision systems rely on very large numbers of training images to learn to recognize particular shapes and configurations of shapes. Traditionally, such datasets of training images needed to be selected and tagged (or labelled) manually. To recognize a particular object, such as a dog or vehicle, under realistic settings with an acceptable degree of reliability, may require data sets of thousands of images per object class. To improve this, a method is provided to generate datasets with a multiplicity of corresponding images are generated using a 3D rendering engine using a plurality of lighting arrangements and a plurality of views. Artefacts may also be introduced. In this way, very large data sets become feasible, with a variable degree of correspondence in each data set.