TL;DR
The authors introduce AdaEye, a systematic framework for scaling down face recognition models specifically for the periocular (eye) region. By optimizing input resolution and backbone architecture, they achieve a 3x increase in throughput and a reduction in computational cost from 6.30 to 1.64 GMACs, while maintaining robust verification accuracy in challenging cross-spectral (Visible to Near-Infrared) scenarios.
What Problem Does This Paper Address?
Full-face recognition is the gold standard for biometric security, but it often fails in the real world. Masks, sunglasses, poor lighting, or privacy-preserving "data minimization" can obscure the lower face. In these cases, the periocular region—the area encompassing the eyes, eyelids, and eyebrows—becomes the most reliable biometric cue.
However, most existing systems simply take a massive full-face model and "mask" the input. This is computationally wasteful. The authors ask: If we only have the eye region, how much can we shrink the neural network without losing accuracy? They specifically focus on "cross-spectral" verification, where a system might compare a high-quality visible-light (VIS) mugshot to a grainy near-infrared (NIR) image captured by a surveillance camera at night.
Methods
The researchers developed AdaEye, adapting the popular AdaFace framework. Instead of proposing a brand-new architecture, they performed a "controlled accuracy-efficiency study" by scaling four key dimensions:
- Input Region: Comparing full-face (112x112), masked-face, and direct periocular crops (80x40).
- Backbone Depth: Switching between heavy (IR-50) and lightweight (IR-18) ResNet-style backbones.
- Channel Width: Using multipliers (1.0, 0.75, 0.5) to reduce the number of filters in each layer.
- Embedding Dimensionality: Compressing the final identity vector from 512 down to 128.
Datasets
The models were trained from scratch on the massive MS1MV2 dataset. For evaluation, the authors used a diverse battery of benchmarks:
- CASIA NIR-VIS 2.0 & ND-NIVL: Used for cross-spectral (VIS-NIR) verification.
- FERET: A standard VIS-VIS face protocol.
- UFPR-Periocular: A "native" periocular dataset captured via mobile devices in unconstrained environments, providing a true test of real-world robustness.
Key Insights and Ideas
- Direct Training Trumps Masking: Simply zeroing out pixels in a full-face model (masking) is less effective than training a model specifically on 80x40 periocular crops. The "direct" models are better adapted to the reduced visible region.
- The Width vs. Depth Trade-off: Reducing the "width" (number of channels) of the network provides a much smoother accuracy-efficiency trade-off than simply jumping to a shallower "depth" (fewer layers).
- Numerical Instability at the Edge: The authors discovered a "convergence boundary." When you combine aggressive backbone reduction with small embedding sizes (e.g., 128-dim), training becomes numerically unstable and prone to failure across different random seeds.
- Protocol Sensitivity: A model that excels on face-derived crops might struggle on native periocular data (like UFPR). This suggests that "periocular recognition" isn't a single task, and model selection must be tailored to how the data is acquired.
Key Results
- Efficiency Gains: Moving from a full-face IR-50 model to a dedicated 80x40 periocular model reduced the computational load from 6.30 GMACs to 1.64 GMACs per image.
- Speedup: On an NVIDIA H100, inference throughput increased by approximately three times.
- Biometric Accuracy: On the UFPR dataset, the dedicated periocular baseline actually outperformed the full-face reference, proving that specialized models can be both faster and more accurate for specific regions.
- Template Size: Reducing embedding dimensionality significantly lowered storage requirements—a critical factor for edge devices with limited memory—with only marginal hits to accuracy in VIS-NIR settings.
Limitations and Future Work
While the study is comprehensive, the authors note that their training protocol was fixed. It is possible that smaller models could perform even better with specialized hyperparameters or distillation techniques. Additionally, while throughput was measured on high-end server GPUs, the real-world impact on mobile or "edge" hardware latency remains an area for future empirical testing. The authors have committed to making all models available on GitHub to facilitate further research in hardware-restricted biometrics.
Summary updated