Exploring how weak supervision can assist the annotation of computer vision datasets : enhancing image annotation workflows via CAMs

Abstract

Current artificial intelligence (AI) workflows depend on researchers performing laborious annotation work. In the case of computer vision (CV), annotators would need to produce bounding boxes and labels in tasks requiring localisation. Since this process needs to be iterated over thousands of images, this can possibly cause low-quality annotations to emerge. Techniques like crowdsourcing can mitigate the effects of this issue, but can still possibly yield a sub-par dataset due to the inexperience of the annotators on either AI or the application itself. However, through existing weakly supervised frameworks that utilise techniques like heuristic rules and other notable methods, the general public can provide even more trustworthy annotations. This paper proposes a basis for an image dataset annotator helper that combines a weakly supervised technique known as class activation maps (CAMs) with convolutional neural networks (CNNs). This produces weakly supervised object localisers that could further improve human image annotation performance. In addition to this, surveys were carried out to extract information that helped either create primary data or verify the results further. Comparing these models with primary crowdsourcing data revealed that the models can annotate better than humans by 9.7% when measuring the localisation error (LE) while taking into account both false positives (FPs) and false negatives (FNs). Moreover, the models can also save up to 36% of the time required to perform manual image annotation. This confirms that there is potential within CAM-empowered models to further improve the image annotation experience.