Platypus Detection Through Deep Learning

Camera traps are essential tools for ecological research. They can be used to directly collect images of wildlife, however, often such images are poorly lit, do not contain the object of interest and have occlusion challenges. The platypus represents one such challenging scenario. To study platypus populations, ecologists face challenges of filtering large numbers of images, most of which do not contain the target species. To solve this problem, in this paper we explore two comprehensive datasets of platypus, one based on automated data augmentation and the other without augmentation. The data sets are based on images captured using camera traps by ecology researchers at the University of Melbourne. Four state-of-the-art object detection models are used and find-tuned on the datasets: YOLOv8, GELAN, Faster R-CNN and Sparse R-CNN. Experimental results for the fine-tuned models show a mean average precision of 0.8 with the YOLOv8 model. This work provides a step change in how ecologists can assess images from camera traps without recourse to manual classification.