Active learning for visual wildlife monitoring

Abstract

Wildlife monitoring and tracking are vital for the preservation of various wildlife species. Activities such as poaching and loss of habitat have made wildlife monitoring and tracking increasingly tricky. Modern-day computer vision techniques have allowed for the collection of data, widely through camera trap images and even UAVs. This work targets the use of camera trap images for re-identification of various wildlife species. To account for the vast pose variations that occur in these camera trap images, we look to make the keypoint estimation mechanism more robust to such pose variations by leveraging common features across multiple images.