Object detection for Verification Based Annotation

Abstract

In this paper, we examine the properties of an object detector for Verification Based Annotation (VBA), where annotation is performed by having machine annotations checked and verified (and corrected, in this work) by a human annotator. We show that with a few modifications from standard practice, a convolutional neural network based object detector can be a robust aid to annotation, without a large degree of parameter tweaking. We previously annotated a variety of small-scale datasets which we attempt to validate some of our ideas and assumptions for an object detector used for VBA.Our approach is to use high-resolution images, training on image crops (as opposed to the usual practice of resizing input images to a fixed resolution), and find this method successful, being more accurate and robust than down-scaling. A particular interest is the effect of localisation noise and systematic bias in annotations. We characterise the impact on object detection performance and compare to human levels, we find noise has a large impact, especially with fewer training examples.