Towards automatic wild animal monitoring: identification of animal species in camera-trap images using deep CNNs

Gomez Villa and colleagues were among the early adopters of very deep convolutional neural networks for camera-trap species identification, testing architectures on challenging real-world images from the Snapshot Serengeti dataset. They documented how factors such as image quality, occlusion, lighting, and the prevalence of empty frames affect classification accuracy, and showed that deeper networks improved performance. The study helped establish that modern deep-learning architectures could handle the messy, variable conditions of camera-trap imagery, motivating wider adoption. It also highlighted the difficulty posed by imbalanced datasets dominated by common species and empty images. As an early demonstration of deep CNNs on camera data, the paper is frequently cited in the development of automated wildlife image classification.