Advances in automated wildlife photo-identification

Abstract

The wild species photo-identification is a basic resource for obtaining the necessary information for several biological research tasks. Today crowdsourcing and citizen science are beginning to play an important role in collecting scientific data. This data source makes it possible to considerably increase the number of records in the sampling database for different scientific projects, especially those related to photographic capture-recapture models of wildlife. However, while the amount of data collected from non-scientific sources increases, a new challenge is presented, mass processing in an agile and efficient way, which allows cleaning and selecting the relevant data for the next stages.
This work addresses the automation of the first stage of the cetacean photo-identification process, which is the detection of the presence or absence of the region of interest in the image (ROI). For this aim, a general-purpose convolutional neural network (Mask R-CNN) was specialized with dolphins images of Cephalorhynchus commersonii specie, collected at different sites on the Patagonian coast over a period of seven y