Extracting identifying contours for African elephants and humpback whales using a learned appearance model

Perspectives in machine learning for wildlife conservation

“…Wildbook supports collaborative mark-recapture, molecular ecology, and social ecology studies, especially where community science and artificial intelligence can help scale-up projects. The image analysis of Wildbook can start with images from any source-scientists, camera traps, drones, community scientists, or social media-and use ML and computer vision to detect multiple animals in the images 100 to not only classify their species, but identify individual animals applying a suite of different algorithms 101,147 . Wildbook provides a technical solution for wildlife research and management projects for noninvasive individual animal tracking, population censusing, behavioral and social population studies, community engagement in science, and building a collaborative research network for global species.…”

“…Identifying individuals from images is even more challenging than species recognition, since the distinctive body patterns of individuals might be subtle or not be sufficiently visible due to occlusion, motion blur, or overhead viewpoint in the case of aerial imagery. Yet, conventional 101 and more recently DL-based 38,54,102 methods have reached strong performance for specific taxa, especially across small populations. Some species have individually-unique coat or skin markings that assist with re-identification: for example, accuracy exceeded 90% in a study of 92 tigers across 8000 video clips 103 .…”

Comparison of Three Individual Identification Algorithms for Sperm Whales (Physeter macrocephalus) after Automated Detection

“…With a machine learning detector trained and configured to extract sperm whale fluke annotations, we then used those annotations and related metadata (in particular the known identifications of the flukes) in the WBIA pipeline (Parham et al 2018) to first custom train the Pose Invariant Embeddings (PIE) algorithm (Moskvyak et al 2019) and then evaluate a) the custom-trained PIE model, b) the pre-trained CurvRank v2 algorithm (Weideman et al 2020; trained on humpback flukes), and c) the pre-trained OC/DTW (Jablons et al 2016) algorithms. The CurvRank v2 and OC/DTW algorithms had already been deployed for sperm whale fluke matching in Flukebook.org and shown anecdotally to be able to match individuals based on their independent extraction of the trailing edge of the fluke.…”

“…The Flukebook.org online platform (Blount et al submitted; Flukebook 2021) provides a Web-based data management framework and a computer vision pipeline (Parham et al 2018) for detection and individual identification of multiple species of cetaceans, including sperm whales. However, existing fluke matching techniques deployed on Flukebook (Jablons 2016; Weideman et al 2020) have not yet been evaluated specifically on sperm whales, leaving questions about their accuracy and reliability for this species. Additionally, new developments in machine learning have suggested that a new class of deep learning-based algorithms may offer an improvement in automated matching capability (Capgemini 2020).…”

“…Here, we test three algorithms applied to this species for the first time. Two of the algorithms were Oriented Curve/Dynamic Time Warping (OC/DTW) (Jablons 2016), and CurvRank v2 (Weideman et al 2020) both of which were off-the-shelf models previously trained on humpback whale ( Megaptera novaeangliae ) flukes. The third is called Pose Invariant Embeddings (PIE v2; Moskvyak et al 2019), which was first tested on manta ray bellies ( Mobula spp.)…”