Synthesising unseen image conditions to enhance classification accuracy for sparse datasets: Applied to chimpanzee face recognition

To aid automated non-invasive population monitoring, we explore chimpanzee face recognition accuracy using a number of algorithms on images with pose and illumination variation, by synthesising images from a generic 3D model. The expense of expeditions and uncontrollable nature of this wild species and environment requires automated face recognition techniques to be robust to pose and illumination variance without incurring additional data collection or manual annotation costs. Unlike for humans, prior knowledge of chimpanzee face shape does not exist, leading us to synthesise 2D images from a custom-built generic 3D shape model for augmenting training and testing data. We use the resulting synthesised images to profile five existing face recognition algorithms. We show that synthetic data can be used to constructively augment training data, as three recognition algorithms have significantly increased accuracy for pose-offset data when augmenting the training data as compared to real data alone.