Object Re-Identification with Synthetic Training Data in Industrial Environments

Abstract

Traditional object detection methods require large amounts of training data. Eliminating the need to generate this data saves time and manual or computational effort. In the field of object re-identification (Re-ID) the similarity of objects is defined by comparing query and target images. Given a query image from an object, these algorithms are able to retrieve the object in a target image, even if the class of the object is not known to the network. In industrial environments, CAD data usually exists for mechanical components. We present a system that uses this CAD data to generate query images to retrieve mechanical components in real images. We train our system exclusively with synthetic training data and evaluate the reality gap when testing real target images. In the evaluation we show, that our system is also able to retrieve unknown classes it has not been trained with. Finally, we identify improvements of our system and address potential research directions of object Re-ID in the future.