摘要
本文提出了一种基于粗到精级联回归树（Ensemble of Regression Trees, ERT）的实时人脸关键点回归方法——DCF E。我们采用一个简单的卷积神经网络（Convolutional Neural Network, CNN）生成关键点位置的概率图，随后通过ERT回归器对这些概率图进行精细化优化。ERT回归器的初始化基于将三维人脸模型拟合到关键点概率图上。ERT的粗到精结构有效缓解了部件形变带来的组合爆炸问题。同时，借助三维人脸模型，我们还解决了回归器初始化的鲁棒性、自遮挡以及正面与侧脸图像的同步分析等关键挑战。在实验中，DCF E在AFLW、COFW以及300W的私有与公开数据集上均取得了目前报道的最佳性能。
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| face-alignment-on-300w | DCFE | NME_inter-ocular (%, Challenge): 5.22 NME_inter-ocular (%, Common): 2.76 NME_inter-ocular (%, Full): 3.24 NME_inter-pupil (%, Challenge): 7.54 NME_inter-pupil (%, Common): 3.83 NME_inter-pupil (%, Full): 4.55 |
| face-alignment-on-300w-split-2 | DCFE | AUC@8 (inter-ocular): 52.42 FR@8 (inter-ocular): 1.83 NME (inter-ocular): 3.88 |
| face-alignment-on-cofw | DCFE | NME (inter-pupil): 5.27% |
| face-alignment-on-ibug | DCFE (inter pupils normalization) | Mean Error Rate: 7.54% |
| facial-landmark-detection-on-300w | DCFE (Inter-ocular Norm) | NME: 3.24 |
| facial-landmark-detection-on-aflw-full | DCFE (Box height Norm, 19 landmarks - no earlobs) | Mean NME : 2.17 |