Weakly semi-supervised object detection for annotation efficiency

Abstract

Annotating datasets is a common obstacle for many industries that may limit the potential for adopting machine learning methods. One example of such an industry is agriculture. Resources may be limited, especially in developing areas, but there is great potential for machine learning models to be used for applications such as tracking diseases. This work revolves around developing an efficient machine learn ing (ML) pipeline and using it to detect coffee berry disease (CBD) in a dataset with images of coffee plants. CBD is a fungal plant pathogen that is difficult to manage and often causes major problems for coffee production. Three common methods to alleviate the burden of manually annotating datasets are semi-supervised learning, weak supervision, and utilizing machine learning in the labelling process. Recently developed open-set detectors that boast impressive performance have a natural use case in this process. These models can predict bounding boxes for arbitrary objects without specific training and can therefore be used to generate proposals for ground truth bounding boxes in a dataset. Following this initial step, manual annotation using time-efficient point labels for the remaining objects in each image results in a mix of strong box labels and weak point labels in each image. This work explores this setting and proposes two models for the task; Point-guided loss suppression (PLS) and mixed Point-Teaching (MPT). The PLS model is a simple adaptation of YOLOv8, which when compared to the semi-supervised case gives a slight improve ment in performance on the CBD dataset and a slight decrease in the MS COCO benchmark. The MPT framework consists of two models where one model is used to generate boxes that the other model uses as pseudo labels during training. The resulting performance for the MPT framework is generally worse, only performing above the baseline in a few cases. The exact efficiency of utilizing point labels is dif ficult to determine, but our results indicate that there are potential use cases where annotating points is more efficient than boxes, especially with further development of the models.