Efficient Camera Trap Image Annotation Using YOLOv5

Abstract

Using camera traps to acquire wildlife images is becoming more common within conservancies. The information provided by these camera traps enhances understanding of wildlife behaviour and population patterns. The detection and counting of animals present in each of the captured images is valuable information as it can be used to guide conservation efforts. Manual annotation of these wildlife images is a tedious painful process. It is becoming more common to use tools that either use AI to annotate camera trap datasets or use AI to aid in annotation. These AI tools are usually trained on species endemic to a particular region. The ability to fine-tune such models to species endemic to one’s particular region is important to save much of the time conservationists manually look through the misclassified images. In this paper, we present a case study where we used a YOLOv5 object detection model trained to detect the presence and count the number of impala and other animals from a dataset collected by researchers at the Dedan Kimathi University of Technology Conservancy. We analyze the results of the AI’s performance with respect to a manually annotated dataset. The model was able to annotate 72% of the dataset at a human level of accuracy. The work here shows promise with regard to time spent labelling camera trap images by leveraging the presence of particular species to autoannotate a majority of the dataset