Abstract
Camera traps are a strategy for monitoring wildlife that collects a large number of pictures. The number of images collected from each species usually follows a long-tail distribution, i.e., a few classes have a large number of instances while a lot of species have just a small percentage. Although in most cases these rare species are the classes of interest to ecologists, they are often neglected when using deep learning models because these models require a large number of images for the training. In this work, we systematically evaluate recently proposed techniques - namely, square-root re-sampling, class-balanced focal loss, and balanced group softmax - to address the long-tail visual recognition of animal species in camera trap images. To achieve a more general conclusion, we evaluated the selected methods on four families of computer vision models (ResNet, MobileNetV3, EfficientNetV2, and Swin Transformer) and four camera trap datasets with different characteristics. Initially, we prepared a robust baseline with the most recent training tricks and then we applied the methods for improving long-tail recognition. Our experiments show that the Swin transformer can reach high performance for rare classes without applying any additional method for handling imbalance, with an overall accuracy of 88.76% for WCS dataset and 94.97% for Snapshot Serengeti, considering a location-based train/test split. In general, the square-root sampling was the method that most improved the performance for minority classes by around 10%, but at the cost of reducing the majority classes accuracy at least 4%. These results motivated us to propose a simple and effective approach using an ensemble combining square-root sampling and the baseline. The proposed approach achieved the best trade-off between the performance of the tail class and the cost of the head classes' accuracy.
Abstract (translated)
URL
https://arxiv.org/abs/2206.12458
