Abstract
Monitoring endangered primate populations is essential for conservation, yet traditional methods are often invasive or prohibitively expensive. This paper discusses PrimNet, a lightweight convolutional neural network (CNN) designed for individual primate identification. By optimizing for small datasets and mobile deployment, PrimNet achieves high accuracy across multiple species while maintaining a footprint small enough for real-time use on standard smartphones.
The Challenge of Primate Tracking
Conservationists and biologists must track individual animals to understand population dynamics and combat illegal trafficking. Historically, this has required physical tagging or GPS collars. These methods are not only costly—ranging from $400 to $4,000 per device—but also invasive. Capturing wild animals for tagging causes significant stress and carries risks of injury or mortality.
While automated facial recognition (FR) is highly effective for humans, applying it to primates in the wild presents unique hurdles. Primate datasets are typically small, and wild environments introduce "unconstrained" variables such as fluctuating light, foliage obstruction, and extreme poses. Standard human FR models, like FaceNet, often overfit when applied to these specialized, low-volume datasets.
Methods: The PrimNet Architecture
To address the "small data" regime, the researchers developed PrimNet, a modified architecture based on SphereNet-4. The system incorporates several key optimizations:
- Group Convolutions: Inspired by ShuffleNet, these make the network "sparser" by limiting channel interactions within filters. This acts as a form of regularization, reducing the parameter count from 12.6 million to under 1 million and preventing the model from simply memorizing training images.
- Width Optimization: To maintain discriminative power despite fewer parameters, the network's hidden layers were made wider (increasing channel counts).
- AM-Softmax Loss: The use of Additive Margin Softmax helps create tight feature clusters for individual identities in high-dimensional space, improving verification accuracy.
- Face Alignment: Images are normalized using a similarity transform based on three manually annotated landmarks: the left eye, right eye, and mouth center.
Datasets and Environmental Complexity
The study utilized three distinct datasets comprising 11,637 images of 280 individuals across 14 species:
- LemurFace: 3,000 images of 129 individuals (Duke Lemur Center).
- GoldenMonkeyFace: 1,450 images of 49 individuals (Rwanda).
- ChimpFace: 5,559 images of 90 individuals (C-Zoo and C-Tai datasets).
Key Results and Insights
PrimNet consistently outperformed state-of-the-art human FR models and previous baseline systems. Key performance metrics include:
- Lemurs: 93.76% Rank-1 identification accuracy.
- Golden Monkeys: 90.36% Rank-1 accuracy.
- Chimpanzees: 75.82% Rank-1 accuracy. The lower performance for chimpanzees is attributed to darker fur and more extreme pose variations in the wild.
Beyond accuracy, PrimNet’s efficiency is a major breakthrough. At only 3.9 MB, it is 10x to 20x smaller than comparable human models. With an inference speed of ~24ms per image, the system was successfully deployed as an Android application, enabling researchers to perform biometric tracking directly in the field.
Conclusion and Future Work
PrimNet demonstrates that specialized architectural priors can overcome the limitations of small datasets in wildlife biometrics. However, the current reliance on manual face alignment remains a bottleneck. Future research will focus on developing robust, automated primate face detectors capable of identifying subjects amidst dense jungle foliage, as well as refining models to better handle the complex facial expressions of great apes.
Summary updated