Semi-Supervised Object Detection for Agriculture

Gabriel Tseng, Krisztina Sinkovics, Tom Watsham, David Rolnick, Thomas C. Walters

Keywords: object detection, dataset, computer vision, semi supervised

TL;DR: A dataset, method and python package for semi-supervised object detection in agricultural contexts

Abstract: Machine learning is a key component of precision agriculture, by allowing plant-level insights to be inferred at scale. However, the labelled data necessary to train these algorithms is expensive to acquire, making methods that leverage unlabelled data -- such as semi-supervised object detection (SSOD) -- of particular interest. Current SSOD methods have been designed specifically for Flickr-based datasets and may not be appropriate for the unique challenges encountered in agricultural contexts, limiting their usefulness in practice. In this paper, we offer innovations in designing, testing, and deploying SSOD in the real world. We compile a challenging new dataset of semi-supervised agricultural images, on which existing SSOD methods largely fail. We present a Python package to more easily test SSOD methods in such real-world domains. Finally, we introduce two components to the standard SSOD pipeline which demonstrably improve performance on our dataset. All code is available at https://github.com/SmallRobotCompany.