Exploring Semi-Supervised Learning for Camera Trap Images from the Wild

Abstract

Camera traps are an important tool for ecologists in their fight against ever increasing animal extinction. However, the use of these camera traps involves the tedious process of manually labeling the animals in captured images. An added hinderance is that of empty images triggered by wind movement and other stimuli called ghost images. Deep learning techniques have previously been applied to automate this task but have been prevented from being entirely effective due to two problems. Firstly, a lack of labeled data due to the expertise of ecologists being required to perform the labeling and secondly the training data being imbalanced in nature due to the high presence of ghost images and images of common animals. Many semi-supervised learning (SSL) algorithms perform well using very small amount of labelled data however need to be evaluated when trained with imbalance data. This paper explores the performance of FixMatch and a derivative called the Auxiliary Balanced Classifier (ABC) under a variety of data imbalance and proportions of labelled data. The algorithms were evaluated using a in the wild imbalanced dataset from camera traps in addition to benchmark datasets such as CIFAR-10, CIFAR-100 and SVHN. While FixMatch showed a consistent drop in performance as the data imbalance was increased, the algorithm generally outperformed ABC. However, the ABC derivative performed better than FixMatch in cases of very high imbalance.