Source-linked AI summary
CurricularFace: Adaptive Curriculum Learning Loss for Deep Face Recognition
Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, Feiyue Huang
TL;DR
Existing face-recognition losses either fail to exploit sample difficultness or emphasize hard samples too early, which can create convergence issues. CurricularFace embeds adaptive curriculum learning into the loss, emphasizing easy samples early and hard samples later with importance determined by stage and difficultness. Experiments on popular facial benchmarks demonstrate superiority over state-of-the-art competitors.
Problem
Existing margin-based methods do not explicitly exploit sample importance, while mining-based methods may emphasize hard samples early and hinder convergence.
Method
CurricularFace adaptively adjusts easy- and hard-sample importance across training stages and assigns hard-sample modulation according to difficultness.
Results
CurricularFace demonstrates superiority over state-of-the-art competitors on popular facial benchmarks.
Takeaways & Limitations
The loss provides an adaptive curriculum that emphasizes easy samples early and hard samples later without manually tuning an additional hyper-parameter.
Abstract
from arXiv · showhide
As an emerging topic in face recognition, designing margin-based loss functions can increase the feature margin between different classes for enhanced discriminability. More recently, the idea of mining-based strategies is adopted to emphasize the misclassified samples, achieving promising results. However, during the entire training process, the prior methods either do not explicitly emphasize the sample based on its importance that renders the hard samples not fully exploited; or explicitly emphasize the effects of semi-hard/hard samples even at the early training stage that may lead to convergence issue. In this work, we propose a novel Adaptive Curriculum Learning loss (CurricularFace) that embeds the idea of curriculum learning into the loss function to achieve a novel training strategy for deep face recognition, which mainly addresses easy samples in the early training stage and hard ones in the later stage. Specifically, our CurricularFace adaptively adjusts the relative importance of easy and hard samples during different training stages. In each stage, different samples are assigned with different importance according to their corresponding difficultness. Extensive experimental results on popular benchmarks demonstrate the superiority of our CurricularFace over the state-of-the-art competitors.
1. Introduction
CurricularFace addresses limitations in margin- and mining-based face-recognition losses by adaptively emphasizing easy samples early and hard samples later. Its modulation function uses training-stage information and sample difficultness, with experiments reporting superiority over state-of-the-art competitors.
- Motivation: Margin-based losses do not exploit sample difficultness, while mining-based losses can overemphasize hard samples early and hinder convergence.ArcFace keeps negative-similarity modulation fixed, whereas MV-Arc-Softmax emphasizes hard samples using a preset constant.
- Method: The method constructs its curriculum online from randomly selected mini-batch samples by mining hard examples and assigning importance according to training stage and difficultness.The relative importance of easy and hard samples changes across training stages, while each hard sample’s importance depends on its own difficultness.
- Contribution: CurricularFace introduces adaptive curriculum learning for face recognition by emphasizing easy samples first and hard samples later.The method is presented as the first work, to the authors’ knowledge, to introduce adaptive curriculum learning for face recognition.
- Method: Its modulation coefficient connects positive and negative cosine similarities without manually tuning an additional hyper-parameter.The coefficient is designed to support adaptive curriculum learning throughout training.
- Experiments: Extensive experiments on popular facial benchmarks demonstrate CurricularFace’s superiority over state-of-the-art competitors.
2. Related Work
Related work combines margin-based feature discrimination with mining-based emphasis on difficult samples, while curriculum learning orders training from easier to harder examples. CurricularFace differs by adapting both the curriculum and hard-sample importance during training.
- Margin-based loss function: Margin-based losses improve discriminability but do not consider each sample’s difficultness.CurricularFace instead emphasizes easy samples first and hard samples later.
- Mining-based loss function: OHEM selects large-loss samples and discards easy samples, whereas Focal loss reduces easy-sample weights using two hyper-parameters.
- Curriculum Learning: Curriculum Learning trains on easier samples before harder ones, but defining sample difficultness with an ad-hoc curriculum can be difficult across problems.Self-Paced Learning emphasizes examples with lower losses as easier samples.
- Curriculum Learning: CurricularFace adaptively changes stage-wise emphasis and assigns different modulation coefficients to hard samples according to their current difficultness.
3. The Proposed CurricularFace
CurricularFace integrates adaptive curriculum learning into face-recognition loss by modulating negative similarities according to sample difficulty and training stage. It suppresses hard samples early, emphasizes them later, and adaptively changes the decision boundary during training.
- The Proposed CurricularFace: CurricularFace incorporates curriculum learning into face recognition by adaptively weighting easy and hard samples across training stages.The method uses online hard-sample mining rather than a fixed sample ordering.
- The Proposed CurricularFace: For misclassified samples, CurricularFace modulates negative cosine similarity with N(t, cosθj) = (t + cosθj) cosθj, while easy samples retain the original similarity.The positive branch adopts ArcFace through T(cos θyi) = cos(θyi + m).
- The Proposed CurricularFace: Hard-sample gradient coefficients M(·) = 2 cosθj + t increase with both negative cosine similarity and the adaptive parameter t.Thus, harder samples receive different importance instead of a single fixed coefficient.
- The Proposed CurricularFace: The parameter t is estimated from the moving average of positive cosine similarities, using EMA to make hard-sample modulation adaptive to the current training stage.The method sets the momentum parameter α to 0.99 and avoids manual tuning of t.
- The Proposed CurricularFace: CurricularFace suppresses hard samples early and emphasizes them later, while its loss-to-ArcFace ratio increases with cosθj as sample difficulty rises.A perceptually good image can still appear hard early when its negative cosine similarities are large.
- Comparison with SOTA Loss Functions: Compared with ArcFace and MV-Arc-Softmax, CurricularFace changes the hard-sample decision boundary across training rather than applying a fixed negative-similarity modulation.The paper also contrasts its adaptive hard-sample weighting with Focal Loss's manually tuned factors.
4. Experiments
Experiments evaluate CurricularFace across standard face-recognition benchmarks, convergence settings, and computational cost. The method generally matches or surpasses strong baselines, while improving convergence with a small backbone.
- Experimental setup: Experiments use CASIA-WebFace and refined MS1MV2 for training, with evaluation on LFW, CFP-FP, CPLFW, AgeDB, CALFW, IJB-B, IJB-C, and MegaFace.ResNet50 and ResNet100 backbones are used under the stated training settings.
- Ablation study: Adaptive estimation of t is more effective than fixed values because it supports learning easier samples first and harder samples later.The ablation compares fixed values 0, 0.3, 0.7, and 1 with adaptive estimation.
- Ablation study: 99.25% accuracy on LFW is achieved with margin m=0.5 using MobileFaceNet, whereas ArcFace does not converge and reaches NAN loss at about the 2,400th step.With m=0.45, both methods converge, but CurricularFace obtains 99.20% versus ArcFace’s 99.10%.
- Comparisons with SOTA methods: 99.80% is achieved on LFW, while CurricularFace surpasses the compared baselines on CFP-FP, CPLFW, AgeDB, and CALFW.The LFW result is described as comparable because performance is near saturated; the other listed benchmarks favor CurricularFace.
- Comparisons with SOTA methods: CurricularFace achieves the best single-model identification and verification performance under both MegaFace protocols, surpassing CosFace, ArcFace, Adacos, P2SGrad, and PFE.It also outperforms competitors under the refined ArcFace testing protocol and improves over ArcFace in identification and verification.
- Computational cost: 0.378s per iteration is required by CurricularFace versus 0.370s for ArcFace on NVIDIA P40 GPUs, while inference cost remains the same as the backbone.The added training cost comes from adjusting negative cosine similarities for hard samples.
5. Conclusions
The paper proposes CurricularFace as an adaptive curriculum loss for deep face recognition. It emphasizes easy samples early and hard samples later, with experiments supporting effectiveness and convergence robustness.
- 5. Conclusions: CurricularFace embeds adaptive curriculum learning into deep face recognition.The paper presents this as an adaptive training strategy rather than a fixed curriculum.
- 5. Conclusions: The method addresses easy samples in early training stages and hard samples in later stages.This training order is the central design principle described in the conclusion.
- 5. Conclusions: Extensive experiments on popular facial benchmarks demonstrate CurricularFace’s effectiveness compared with state-of-the-art competitors.The conclusion also characterizes the method as easy to implement and robust to converge.