Abstract
Researches using margin-based comparison loss demonstrate the effectiveness of penalizing the distance between face features and their corresponding class centers. Despite their popularity, they do not explicitly encourage generic embedding learning for open-set recognition. We analyze margin-based softmax loss from a probabilistic perspective and propose two general principles: monotonic decreasing and margin probability penalty. We introduce "LinearFace" and a new architecture, Linear-Auto-TS-Encoder (LATSE), which treats face recognition as an information transmission problem. Extensive experiments show LATSE achieves over 99% accuracy on the MegaFace benchmark.
Introduction: The Open-Set Challenge
Most modern face recognition systems use "margin-based" loss functions (like ArcFace or CosFace) to push features of the same identity closer together. However, the authors identify two major gaps in current approaches:
- The Open-Set Bottleneck: These systems are often optimized solely for discriminative power (telling A apart from B). They don't explicitly encourage the model to learn a "generic" representation that generalizes well to billions of people not present in the training set.
- Loss Function Instability: Popular loss functions aren't always mathematically "well-behaved." For instance, some aren't monotonically decreasing, which can lead to training instability or sub-optimal convergence when the angle between a face feature and its class center is large.
Methodology: The LATSE Framework
The authors propose a framework called LATSE (Linear-Auto-TS-Encoder), which rests on three pillars designed to improve generalization through "information transmission."
1. A Probabilistic View of Margin Loss
By analyzing margin losses through the lens of probability, the authors establish that the transformation from angle to probability must be monotonically decreasing and include a non-negative margin penalty. Based on this, they introduce LinearFace, a loss function that provides a stable, linear penalty curve.
As shown in the figure above, while popular losses like ArcFace can shrink the margin when the angle is too large, the proposed LinearFace maintains a steady pressure, making it easier for the model to learn during the early stages of training.
2. Generative Supervision and Teacher-Student Learning
The LATSE architecture integrates a generative task and a noise-filtering mechanism to ensure the embedding transmits high-quality identity information.
- Generative Multi-Modal Supervision: Instead of just a classifier, LATSE uses an Auto-Encoder. The encoder creates the face embedding, while a decoder tries to reconstruct a "frontalized" version of the face. This forces the embedding to "transmit" as much identity-relevant information as possible.
- Teacher-Student (TS) Learning: Large-scale face datasets are notoriously noisy. LATSE uses a pre-trained "Teacher" network to filter out noisy gradients in real-time, ensuring the "Student" network only learns from clean, reliable data.
Experimental Results
The models were trained on CASIA-WebFace and MS1MV2/MS1M-RetinaV and evaluated across several industry-standard benchmarks.
- MegaFace Milestone: The proposed LATSE reached 99.14% identification accuracy on MegaFace (with 1 million distractors), outperforming previous state-of-the-art methods like ArcFace and CosFace.
- Robustness to Variation: On datasets with high pose and age variation (CALFW and CPLFW), the method improved accuracy by roughly 1% over standard margin-based approaches, proving better generalization.
- Ablation Success: The authors showed a clear "stacking" effect: adding the Teacher-Student strategy improved results, and adding the Generative (decoder) task improved them even further.
Conclusion
The core insight of this work is that treating face recognition as an information transmission problem leads to more robust embeddings. By ensuring the network can rebuild a face from its embedding, the model learns a richer set of features that are more likely to work on unseen identities. While the LATSE framework is more complex than standard classifiers, its superior performance on open-set benchmarks demonstrates the value of generative supervision in deep face recognition.
Summary updated