AdaSin: Enhancing Hard Sample Metrics with Dual Adaptive Penalty for Face Recognition

TL;DR

AdaSin tackles the core challenge of hard samples in face recognition by introducing a sine-based difficulty measure $D(\theta) = \sin\left(\frac{\theta}{2}\right)$ and a dual adaptive penalty that jointly modulates the positive angular margin and the negative cosine similarities. The modulation coefficient $\\Phi = t^{(k)} + h \, D(\theta_{y_i})$ evolves with training via an EMA-based $t^{(k)}$, enabling curriculum-like emphasis on hard samples in later stages. This approach yields stronger intra-class compactness and inter-class separability, demonstrated through extensive experiments on eight benchmarks, with notable gains on AgeDB-30, CPLFW, and IJB-B/C. Overall, AdaSin provides a flexible, curriculum-guided framework that surpasses several state-of-the-art losses in challenging face-recognition settings.

Abstract

In recent years, the emergence of deep convolutional neural networks has positioned face recognition as a prominent research focus in computer vision. Traditional loss functions, such as margin-based, hard-sample mining-based, and hybrid approaches, have achieved notable performance improvements, with some leveraging curriculum learning to optimize training. However, these methods often fall short in effectively quantifying the difficulty of hard samples. To address this, we propose Adaptive Sine (AdaSin) loss function, which introduces the sine of the angle between a sample's embedding feature and its ground-truth class center as a novel difficulty metric. This metric enables precise and effective penalization of hard samples. By incorporating curriculum learning, the model dynamically adjusts classification boundaries across different training stages. Unlike previous adaptive-margin loss functions, AdaSin introduce a dual adaptive penalty, applied to both the positive and negative cosine similarities of hard samples. This design imposes stronger constraints, enhancing intra-class compactness and inter-class separability. The combination of the dual adaptive penalty and curriculum learning is guided by a well-designed difficulty metric. It enables the model to focus more effectively on hard samples in later training stages, and lead to the extraction of highly discriminative face features. Extensive experiments across eight benchmarks demonstrate that AdaSin achieves superior accuracy compared to other state-of-the-art methods.

Figures (9)

• Figure 1: Comparison of the changes in the values of the modulation coefficient function of our design with the values of the original modulation coefficient function.
• Figure 2: The range of values of the two metric functions.
• Figure 3: The figure above illustrates the specific workflow of our AdaSin loss function. First, we determine whether a sample is a hard sample. If the sample is classified as easy, we directly proceed with the softmax computation. If it is identified as a hard sample, we calculate the sine value of the angle between the sample and the positive class center. Next, we compute the modulation coefficient $\Phi$, and finally, we apply this coefficient to adaptively scale the logits.
• Figure 4: Variable range of modulation coefficients function.
• Figure 5: The red solid line, green solid line, and blue solid line represent the trends of the adaptive parameter $t$, the modulation coefficients of CurricularFace, and the modulation coefficients of our AdaSin, respectively, throughout the training process. We can clearly observe that the modulation coefficient curve of CurricularFace (green solid line) remains consistently below the red dashed line throughout the entire training process.