Animal Localization in Camera-Trap Images with Complex Backgrounds
Praneet Singh, Stacy Marie Lindshield, Fengqing Maggie Zhu, Amy R. Reibman · 2020
Motion-sensor camera traps help collect images of animals in the wild without intruding upon their native habitat. To obtain key insights about animal health and population densities, accurate counting, detection and classification of animals is important. Deep convolution neural networks perform well on these tasks when the background is free from dense vegetation, shadows, occlusions and rapid illumination changes. However, when the camera traps are located in regions with extremely complex backgrounds, performance of these models degrades significantly. This is due to the fact that the models learn to focus on aspects of the image that are unrelated to the animals. In this paper, we propose a system based on Robust Principal Component Analysis (Robust PCA) that spatially localizes the animals in the image. This localization can then be integrated into existing models to improve classification and detection accuracy. We demonstrate that our system creates better localizations than those of a pre-trained R3Net.