
摘要
面部对齐算法在不受限制的情况下定位面部图像中的一组标志点。现有的最先进方法通常在存在遮挡、强烈变形、大姿态变化和模棱两可的配置时会失败或精度下降。本文介绍了一种基于粗到精级联回归树集合的鲁棒且高效的面部对齐算法——3DDE。该算法通过将一个三维人脸模型稳健地拟合到卷积神经网络生成的概率图上来初始化。通过这种初始化,我们解决了自遮挡和大范围人脸旋转的问题。此外,回归器隐式地对解施加了一个先验的人脸形状,从而应对遮挡和模棱两可的人脸配置。其粗到精的结构解决了部件变形的组合爆炸问题。实验结果表明,3DDE在300W、COFW、AFLW和WFLW数据集上的表现优于现有最先进方法。最后,我们进行了跨数据集实验,揭示了这些基准测试中存在的显著数据集偏差。
代码仓库
bobetocalo/bobetocalo_eccv18
tf
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| face-alignment-on-300w | 3DDE | NME_inter-ocular (%, Challenge): 4.92 NME_inter-ocular (%, Common): 2.69 NME_inter-ocular (%, Full): 3.13 NME_inter-pupil (%, Challenge): 7.10 NME_inter-pupil (%, Common): 3.73 NME_inter-pupil (%, Full): 4.39 |
| face-alignment-on-300w-split-2 | 3DDE | AUC@8 (inter-ocular): 53.94 FR@8 (inter-ocular): 2.33 NME (inter-ocular): 3.73 |
| face-alignment-on-cofw | 3DDE (Inter-pupil Norm) | NME (inter-pupil): 5.11% Recall at 80% precision (Landmarks Visibility): 63.89 |
| face-alignment-on-wflw | 3DDE | AUC@10 (inter-ocular): 55.44 FR@10 (inter-ocular): 5.04 NME (inter-ocular): 4.68 |
| facial-landmark-detection-on-300w | 3DDE (Inter-ocular Norm) | NME: 3.13 |
| facial-landmark-detection-on-aflw-full | 3DDE (Box height Norm, 19 landmarks - no earlobs) | Mean NME: 2.01 |