NOSpcimen: A First Approach to Unsupervised Discarding of Empty Photo Trap Images

A key tool in wildlife conservation is the observation and monitoring of wildlife using photo-trapping cameras. Every year, thousands of cameras around the world take millions of images. A large proportion of these are empty - they do not show any animals. Sorting out these blank images requires considerable effort from biologists, who spend hours on the task. It is therefore of particular interest to automate this task. So far, systems have been proposed which are based on the use of supervised learning models. In order to learn, these systems require the annotation of images to indicate where animals are located within them. NOSpcimen (NOn-SuPervised disCardIng of eMpty images based on autoENcoders) system takes a different approach. It relies on unsupervised learning mechanisms. Thus, no prior annotation work is required to automate the process of discarding empty images.