Exploratory Investigation on a Naive Pseudo-labelling Technique for Liquid Droplet Images Detection using Semi-supervised Learning

Annotation are essential in using machine learning for object detection in images. However, good detection will require high quality annotations for supervised learning-based techniques. It is also known that manual annotation by human experts are laborious, tedious and error-prone. This problem is more so challenging in physical sciences images, such as liquid droplets or crystals. In this paper, we suggest a minimal effort approach to train YOLOv3 models using a relatively smaller set of training data for liquid droplet images. The proposed approach uses only minimal amount of annotations or annotated images, and naively build training data set using YOLOv3 models. The investigation compares two approach variants and discuss the results as an exploratory work in finding suitable semi-automatic image annotation techniques for object detection in liquid droplet images.