WildAIID is an umbrella application that combines various third party software to allow users to individually identify animals from a library of images or video using AI. WildAIID has been optimised for the individual identification of eastern grey kangaroos. It should nevertheless be useful for other species of macropods and potentially other mammalian species as well.
WildAIID © 2025 by UNSW Sydney is licensed under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International
WildAIID was created by Tariq Khan, Hamza Farooq, Erik Meijering and Terry J. Ord and should be cited as:
Khan T, Farooq H, Meijering E and Ord TJ (2025). WildAIID: Wildlife Artificial Intelligence Identification. Version 1.0.0. www.ordlab.unsw.edu.au/WildAIID
A NOTE ON APPLICATION: There are two critical challenges associated with applying AI to identify animals in wild populations:
- CHALLENGE 1: Wild populations present an ‘open set’ problem: new individuals are continually coming into a population through birth or immigration, while resident individuals are leaving a population through death or dispersal. The complication here is the AI must be constantly updated (retrained) to accommodate natural recruitment.
- CHALLENGE 2: The AI must be initially trained using tens of thousands of images (if not more) where animals have already been identified by the researcher. The complication here is obtaining these huge ‘training sets’ is often intractable because of the extensive labour required to identify animals in these images.
WildAIID does solve the first challenge by allowing users to continually retrain the existing AI identification system as new individuals come into the population. It still requires a researcher to assign an identity to those new animals, but only these new animals, dramatically cutting workload.
WildAIID does not solve (yet) the second challenge and you will still need a very large number of images to initially train the AI to recognise animals in your population. We are currently adapting the systems underlying WildAIID to allow a more focussed ‘informed’ identification method using a repertoire of salient facial features (distance between eyes, head shape, etc). This will reduce the number of images needed to train the AI and hopefully by a large margin.
All AI systems must cope with these two challenges. AIs that identify the species of animals in images essentially avoid the open-set problem and are pre-trained using millions of images, so you're usually good to go. The individual identification of animals within a population (of the same species) is very different. Be wary of any claim of an AI system with high accuracy (e.g., >95% correct assignment) because this claim will be based on accuracy calculations using a closed set and with an initial training set that is massive.
