Related Experiment Videos
Hybrid CNN with angular margin supervision for robust face identification and verification
Andisani Nemavhola1, Colin Chibaya2, Serestina Viriri3
1School of Consumer Intelligence and Information Systems, University of Johannesburg, Johannesburg, South Africa.
Frontiers in Artificial Intelligence
|August 11, 2026
Summary
This study shows that face recognition system performance varies significantly with supervision methods and datasets. ArcFace does not always outperform Softmax, highlighting the need for careful evaluation across diverse conditions.
Area of Science:
- Computer Science
- Artificial Intelligence
- Biometrics
Background:
- Face recognition systems are crucial for security and identity verification.
- Understanding supervision sensitivity and cross-dataset performance is vital for reliable systems.
Purpose of the Study:
- To investigate supervision sensitivity (Softmax vs. ArcFace) in face recognition architectures.
- To evaluate performance consistency across Labeled Faces in the Wild (LFW) and FAGEv2 datasets.
- To analyze fold-level and cross-dataset performance variations.
Main Methods:
- Evaluated ResNet50, MobileNetV3, DeiT-Small, and a Hybrid architecture.
- Used five-fold subject-disjoint cross-validation on LFW and FAGEv2.
- Measured Top-1 identification accuracy, AUC, EER, and computational complexity.
Main Results:
- Substantial supervision sensitivity observed across architectures and datasets.
- Hybrid-Softmax and DeiT-Small-Softmax showed strong performance, but results varied by dataset and metric.
- ArcFace's impact differed, improving some convolutional models but not transformer or hybrid ones.
Conclusions:
- Additive angular margin supervision (ArcFace) is not universally superior to Softmax.
- Robust face recognition benchmarking requires multi-dataset evaluation, fold-level analysis, and supervision sensitivity assessment.