Facial recognition for lemurs allows researchers to identify them less invasively

Tagging is a stressful experience for lemurs. Now there’s an alternative.

After relying on photographs and their own memories to recognize lemurs in the wild, Rachel Jacobs and Stacey Tecot were frustrated that they couldn’t easily combine their data. Thinking there must be a better way, the primatologists teamed up with their colleague Andrea Baden and a group of computer scientists, among them Anil Jain, to design a facial recognition system for lemurs. LemurFaceID isn’t just good for scientists; it also spares lemurs the stress of being captured and tagged. We spoke with the researchers about their creation.

ResearchGate: Where did this idea come from?

Rachel Jacobs: My dissertation research was a study on color vision in red-bellied lemurs and involved collecting fecal samples from as many individual lemurs as possible. I knew that capturing and collaring wouldn’t be practical, so I worked with a research team in Madagascar to identify individuals based on their facial pelage patterns. We realized they were quite variable, and we could use this variation to identify individuals. But it’s definitely not an easy task, and some people are better at it than others.

Stacey Tecot: When I began my dissertation research, I also needed to recognize individuals to see their response to environmental change. Like Rachel, I had to know exactly which lemurs the behavioral and fecal data were from. My team and I selected groups that already had one individual with a collar and tag from earlier researchers’ work. We used that, as well as facial features and locations, to get a pretty good sense of who individuals were.

Rachel conducted her research on the same population that I studied, and we were really frustrated that we couldn’t combine our datasets, since we named the lemurs differently for our studies. However, Rachel’s superb knowledge of individuals was important for work I did later with Andrea Baden. She was able to tell us who the individuals were. At that point, it really hit home for me: we were having entire conversations about the identity of single individuals, and it was an inefficient method.

Jacobs: Long-term data on known individuals is key to addressing evolutionary questions about this population. Many people need to be trained to accurately identify individuals, and that’s very time-consuming. So we thought, perhaps a computer-assisted recognition program like those used in humans could help. We needed the expertise of computer scientists and reached out to Anil Jain at Michigan State University. Fortunately, he and members of his lab took on the challenge.


“We were having entire conversations about the identity of single individuals.”

RG: How does it work?

Anil Jain: LemurFaceID consists of three stages. The first is preprocessing. The lemur image is processed to locate the eyes, extract the lemur’s face, and normalize the face so that all the lemurs have the same size, orientation and illumination. The next stage is feature extraction; a dataset of lemurs with known identities is used to learn which features will be useful to distinguish different lemurs. The third stage is recognition. Given a lemur with an unknown identity, we process the image, extract the learned features and then compare those features with features of known lemurs in the database.

Tecot: Right now, the system requires that we manually identify the location of each eye, but our next step is to increase the dataset and train the system to automate this process.

RG: How does this improve on existing tracking methods?

Jacobs: A common method of identifying individual lemurs for research involves capturing them and placing unique collars on them. This method has a lot of benefits. For example, you can be fairly confident that different observers are making correct identifications. Also, since you have animals in hand at some point, you can collect data that would otherwise be impossible, like morphometric data, information on ectoparasites, and blood samples. But there are risks to this method. Capturing can result in injury. It can be highly stressful. It can result in changes in group compositions or have adverse health effects. It also isn’t practical for larger-scale studies where you want information on a lot of individuals. LemurFaceID by contrast is relatively non-invasive. There’s no need to capture individuals. And compared to manually identifying individuals by facial patterns it’s significantly less time consuming.

Tecot: That is particularly important when you have several research assistants, students, and local guides. We want to be sure everyone is skilled in identifying individuals, and this tool can enhance that training and double-check identifications throughout the study. If a new animal enters a group, we want to know with certainty if it’s someone we previously knew.


“This research truly represents a cross-disciplinary collaboration.”

RG: Do you anticipate any other uses in the future?

Jacobs: We’d like to see this used by researchers to keep track of lemur group compositions in the wild, so that we can obtain long-term demographic data. This is important both for evolutionary studies, as well as conservation. Down the line, we’d also like to see it used by law enforcement officials, local citizens, and tourists. Because the system makes identifications based on photographs, anyone out there with these wild populations could potentially help us keep track of what’s going on.

Tecot: This could also help people experience a more personal connection with the wildlife they come across when they visit the park. A pie-in-the-sky idea is to also use this technology to determine how individual phenotypes change through time, which would also enhance our ability for long-term research.

RG: Could it be adapted to work for other species?

Jacobs: We certainly think so. Many people might not know it, but there are actually over 100 lemur species in Madagascar. We think this could be useful for many of them. We also think it could be used for non-lemur species that have similar variable facial patterns, such as some bears.

RG: What was it like developing this solution to your problem?

Jacobs: This research truly represents a cross-disciplinary collaboration. Three of the coauthors on this manuscript are lemur biologists, and the remaining coauthors are computer scientists. Knowing very little about face recognition, Stacey and I knew that we had to reach out to experts.

Tecot: The end result demonstrates that individuals from two very different fields of study can successfully work together to address challenges that might otherwise go unaddressed, resulting in an expansion of the research we can conduct in the field.

Featured image courtesy of Tee La Rosa.