Facial recognition AI trained to work on bears
For most people, assessing the weight or color of a bear’s fur is not a top priority during an unexpected encounter in the woods. Instead, the desire to survive usually outweighs lingering to admire the predator’s large claws or snout shape. Knowing this, you would be forgiven for having difficulty telling one bear from another.
For many conservationists, monitoring individual animals over long periods of time, even years, is crucial to conservation efforts. But even experts are easily mistaken. This is especially true given the often dramatic seasonal fluctuations in a bear’s weight, as well as how they can appear physically different before and after hibernation. To help wildlife biologists make sense of all this, a team from EPFL in Switzerland and Alaska Pacific University (APU) developed PoseSwin, a machine learning program that can tell brown bears apart from each other. The technology was recently detailed in a study recently published in the journal Current cell biology.
PoseSwin was trained on more than 72,000 photos of 109 different brown bears taken by APU researcher Beth Rosenberg between 2017 and 2022. Rosenberg captured the images at all times of the day and night and in various weather conditions, while also making sure to document the bears in a variety of behaviors. She and her colleagues then drew on their existing knowledge of brown bear physiology to determine the few anatomical details that remain relatively constant throughout the animal’s life. These characteristics include the angle of the brow bone, ear placement, and muzzle shape. Next, they incorporated data on how the bears looked in different poses and from different angles.
“Our biological intuition was that head features combined with pose would be more reliable than body shape alone, which changes dramatically with weight gain,” explained Alexander Mathis, project collaborator and researcher at EPFL’s Brain Mind Institute and Neuro-X Institute. “The data proved us right: PoseSwin significantly outperformed models that used body images or ignored pose information.”
From there, the team tested PoseSwin in the field with the help of citizen scientists. After collecting more brown bear portraits from visitors to Katmai National Park and Preserve (home of Fat Bear Week), the researchers fed the photos into the machine learning program. In several cases, PoseSwin was able to match individual bears to those already in its database. PoseSwin designers were already able to begin tracking how and where these predators moved in search of seasonal food.
“This is a real-world example of the potential of the PoseSwin model,” Rosenberg said. “The technology could eventually be used to analyze the thousands of photos visitors take each year and help build a map of how brown bears use this vast area.”
Rosenberg and his colleagues are now using PoseSwin to monitor more than 100 bears living around the McNeil River State Game Sanctuary without disrupting their daily habits. In doing so, they should gain more accurate information about the health and welfare of bears, providing a much-needed boost to conservation efforts.
“Bears are at the top of the food chain and keep their ecosystem functioning properly. They are essential to maintaining healthy systems,” Rosenberg explained.
PoseSwin probably won’t stay as bear-centric. Early benchmark tests indicate it is also incredibly accurate when trained on macaques, suggesting it could soon expand to many other species. The machine learning algorithm is also available as open source, so anyone can access it for their own topic, although chances are none of them will be harder for PoseSwin to identify.
“Bears are perhaps the most difficult species to recognize individually,” Mathis said. “We first focused on them with the idea that our program could be adapted to other species, from mice to chimpanzees.”
The best new PopSci 2025 releases
