Exploring the Use of Deep Learning with Crowdsourcing to Annotate Images

We investigate what, if any, benefits arise from employing hybrid algorithm-crowdsourcing approaches over conventional approaches of relying exclusively on algorithms or crowds to annotate images.  We introduce a framework that enables users to investigate different hybrid workflows for three popular image analysis tasks: image classification, object detection, and image captioning.   Three hybrid approaches are included that are based on having workers: (i) verify predicted labels, (ii) correct predicted labels, and (iii) annotate images for which algorithms have low confidence in their predictions.  Deep learning algorithms are employed in these workflows since they offer high performance for image annotation tasks.  Each workflow is evaluated with respect to annotation quality and worker time to completion on images coming from three diverse datasets (i.e., VOC, MSCOCO, VizWiz). Inspired by our findings, we offer recommendations regarding when and how to employ deep learning with crowdsourcing to achieve desired quality and efficiency for image annotation.