A step towards the practical evaluation of wildlife identification algorithms

Abstract

Computer vision approaches have shown to be an effective tool for wildlife identification across a wide range of species. As the field grows and the amount of available data grows with it, deep learning based approaches are beginning to enter as alternatives to previous methods. In this thesis we examine one of these newer deep learning methods --- Pose Invariant Embeddings (PIE) --- to determine how it performs in an environment representative of what we anticipate in practice. This environment is based on the idea of continual curation, where an initially small database evolves and grows over time. We find that the training data required for PIE results in an additional complication that can have a large impact on perceived performance when evaluating individuals seen and unseen during training. We compare PIE against a baseline algorithm HotSpotter, and find that under our current dataset sizes the baseline remains the preferred approach. However, PIE presents a greater opportunity as we move forward with additional data, which can eventually lead to better performance as our continually curated dataset develops.