Location Invariant Animal Recognition UsingMixed Source Datasets and Deep Learning

Abstract

O_LIA time-consuming challenge faced by camera trap practitioners all over the world is the extraction of meaningful data from images to inform ecological management. The primary methods of image processing used by practitioners includes manual analysis and citizen science. An increasingly popular alternative is automated image classification software. However, most automated solutions are not sufficiently robust to be deployed on a large scale. Key challenges include limited access to images for each species and lack of location invariance when transferring models between sites. This prevents optimal use of ecological data and results in significant expenditure of time and resources to annotate and retrain deep learning models. C_LIO_LIIn this study, we aimed to (a) assess the value of publicly available non-iconic FlickR images in the training of deep learning models for camera trap object detection, (b) develop an out-of-the-box location invariant automated camera trap image processing solution for ecologist using deep transfer learning and (c) explore the use of small subsets of camera trap images in optimisation of a FlickR trained deep learning model for high precision ecological object detection. C_LIO_LIWe collected and annotated a dataset of images of "pigs" (Sus scrofa and Phacochoerus africanus) from the consumer image sharing website FlickR. These images were used to achieve transfer learning using a RetinaNet model in the task of object detection. We compared the performance of this model to the performance of models trained on combinations of camera trap images obtained from five different projects, each characterised by 5 different geographical regions. Furthermore, we explored optimisation of the FlickR model via infusion of small subsets of camera trap images to increase robustness in difficult images. C_LIO_LIIn most cases, the mean Average Precision (mAP) of the FlickR trained model when tested on out of sample camera trap sites (67.21-91.92%) was significantly higher than the mAP achieved by models trained on only one geographical location (4.42-90.8%) and rivalled the mAP of models trained on mixed camera trap datasets (68.96-92.75%). The infusion of camera trap images into the FlickR training further improved AP by 5.10-22.32% to 83.60-97.02%. C_LIO_LIEcology researchers can use FlickR images in the training of automated deep learning solutions for camera trap image processing to significantly reduce time and resource expenditure by allowing the development of location invariant, highly robust out-of-the-box solutions. This would allow AI technologies to be deployed on a large scale in ecological applications. C_LI