附录B 外文原文
Look at Boundary: A Boundary-Aware Face Alignment Algorithm
Wenyan (Wayne) Wu 1,2, Chen Qian2, Shuo Yang3, Quan Wang2, Yici Cai1, Qiang Zhou1 1Tsinghua National Laboratory for Information Science and Technology (TNList), Department of Computer Science and Technology, Tsinghua University
2SenseTime Research
3Amazon Rekognition
1wwy15@mails.tsinghua.edu.cn caiyc@mail.tsinghua.edu.cn 1 zhouqiang@tsinghua.edu.cn
2 {qianchen, wangquan}@sensetime.com 3 shuoy@amazon.com
Abstract
We present a novel boundary-aware face alignment algorithm by utilising boundary lines as the geometric structure of a human face to help facial landmark localisation. Unlike the conventional heatmap based method and regression based method, our approach derives face landmarks from boundary lines which remove the ambiguities in the landmark definition. Three questions are explored and answered by this work: 1. Why using boundary? 2. How to use boundary? 3. What is the relationship between boundary estimation and landmarks localisation? Our boundary-aware face alignment algorithm achieves 3.49% mean error on 300-W Fullset, which outperforms state-of-the-art methods by a large margin. Our method can also easily integrate information from other datasets. By utilising boundary in-formation of 300-W dataset, our method achieves 3.92% mean error with 0.39% failure rate on COFW dataset, and 1.25% mean error on AFLW-Full dataset. Moreover, we propose a new dataset WFLW to unify training and testing across different factors, including poses, expressions, illu-minations, makeups, occlusions, and blurriness.
1 Experiments
Datesets. We conduct evaluation on four challenging datasets including 300W [39], COFW [5], AFLW [28] and WFLW which is annotated by ourself.
300W [39] dataset: 300W is currently the most widely-used benchmark dataset. We regard all the training samples (3148 images) as the training set and perform testing on (i) full set and (ii) test set. (i) Full set contains 689 images and is split into common subset (554 images) and challenging subsets (135 images). (ii) Test set is the private test-set used for the 300W competition which contains 600 images.
COFW [5] dataset consists of 1345 images for training and 507 faces for testing which are all occluded to different degrees. Each COFW face originally has 29 manually an-notated landmarks. We also use the test set which has been re-annotated by [19] with 68 landmarks annotation scheme to allow easy comparison to previous methods.
AFLW [28] dataset: AFLW contains 24386 in-the-wild faces with large head pose up to 120◦ for yaw and 90◦ for pitch and roll. We follow [72] to adopt three settings on our experiments: (i) AFLW-Full: 20000 and 4386 images are used for training and testing respectively. (ii) AFLW-Frontal: 1314 images are selected from 4386 testing images for evaluation on frontal faces.
WFLW dataset: In order to facilitate future research of face alignment, we introduce a new facial dataset base on WIDER Face [61] named Wider Facial Landmarks in-the-wild (WFLW), which contains 10000 faces (7500 for training and 2500 for testing) with 98 fully manual annotated landmarks. Apart from landmark annotation, out new dataset includes rich attribute annotations, i.e., occlusion, pose, make-up, illumination, blur and expression for comprehensive analysis of existing algorithms. Compare to previous dataset, faces in the proposed dataset introduce large variations in expression, pose and occlusion. We can simply evaluate the robustness of pose, occlusion, and expression on proposed dataset instead of switching between multiple evaluation protocols in different datasets. The comparison of WFLW with popular benchmarks is illustrated in the supplementary material.
Evaluation metric. We evaluate our algorithm using standard normalised landmarks mean error and Cumulative Errors Distribution (CED) curve. In addition, two further statistics i.e. the area-under-the-curve (AUC) and the failure rate for a maximum error of 0.1 are reported. Because of various profile face on AFLW [28] dataset, we follow [72] to use face size as the normalising factor. For other dataset, we follow MDM [51] and [39] to use outer-eye-corner distance as the “inter-ocular” normalising factor. Specially, to compare with the results that reported to be normalised by “inter-pupil” (eye-centre-distance) distance, we report our results with both two normalising factors on Table 1.
Implementation details. All training images are cropped
|
Method |
Common |
Challenging |
Fullset |
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
Subset |
Subset |
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
Inter-pupil Normalisation |
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
RCPR [6] |
6.18 |
17.26 |
8.35 |
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
CFAN [69] |
5.50 |
16.78 |
<p 剩余内容已隐藏，支付完成后下载完整资料
附录A 译文 看边界：边界感知面部对齐算 文言（韦恩）吴、陈倩、朔阳、全旺、易慈茶、强周 ITSINGHUA信息科学与技术国家实验室（TNLIST） 清华大学计算机科学与技术 感官石灰研究 萨马勋.雷克妮 摘 要 我们通过利用边界线作为人脸的几何结构来提出一种新颖的边界感知人脸对齐算法，以帮助进行面部地标定位。与传统的基于热图的方法和基于回归的方法不同，我们的方法从边界线导出面部地标，这消除了界标定义中的模糊性。这项工作探讨并回答了三个问题：1。为什么要使用边界？2.如何使用边界？3.边界估计与地标定位之间有什么关系？我们的边界感知面部对齐算法在300 W全集上实现了3.49％的平均误差，其优于大多数最先进的方法。我们的方法还可以轻松地整合来自其他数据集的信息。利用300W数据集的边界信息，我们的方法实现了3.92％的平均误差，COFW数据集的失效率为0.39％，AFLW-Full数据集的平均误差为1.25％。此外，我们提出了一个新的数据集WFLW，以统一不同因素的训练和测试，包括姿势，表情，照明，化妆，遮挡和模糊。 1实验 数据集。我们对四个具有挑战性的数据集进行评估，包括300W[39], COFW[5], AFLW[28] 和WFLW，由我们自己注释。 300W [39]数据集：300W是目前使用最广泛的基准数据集。我们将所有训练样本（3148个图像）视为训练集，并对（i）全集和（ii）测试集进行测试。（i）全集包含689个图像，并被分成共同子集（554个图像）和具有挑战性的子集（135个图像）。（ii）测试集是用于300W竞赛的私人测试集，其包含600个图像。 COFW[5]数据集包括1345个用于训练的图像和507个用于测试的面部，它们都被遮挡到不同程度。每个COFW面最初具有29个手动注释的界标。我们还使用了经过重新注释的测试集[19]使用68个标记注释方案，以便与以前的方法轻松比较。 AFLW[28]数据集：AFLW包含24386个野外脸部，大头部姿势高达120◦ 用于偏航，90◦ 用于俯仰和滚转。我们跟着[72]在我们的实验中采用三种设置：（i）AFLW-Full：20000和4386图像分别用于训练和测试。(ii)正面：从4386个测试图像中选择1314个图像，用于在正面评估。 WFLW数据集：为了便于将来研究人脸对齐，我们在WIDER Face上引入了一个新的面部数据集[61]名为Wider Facial Landmarks in the wild（WFLW），包含10000个面（7500个用于训练，2500个用于测试），带有98个完全手动注释的地标。除了地标注释外，新的数据集还包括丰富的属性注释，即遮挡，姿势，化妆，光照，模糊和表达，以便对现有算法进行全面分析。与先前的数据集相比，建议的数据集中的面部在表达，姿势和遮挡方面引入了大的变化。我们可以简单地评估建议数据集上的姿势，遮挡和表达的稳健性，而不是在不同数据集中的多个评估协议之间切换。WFLW与流行基准的比较在补充材料中说明。 评估指标。我们使用标准归一化地标平均误差和累积误差分布（CED）曲线来评估我们的算法。此外，还有两个统计数据，即曲线下面积（AUC）和失败 -报告的最大误差为0.1。由于AFLW上的各种轮廓面[28] 数据集，我们遵循[72] 使用面部大小作为标准化因子。对于其他数据集，我们遵循MDM[51]和[39]使用外眼角距离作为“眼间”归一化因子。特别地，为了与报告通过“inter-pupil”（眼睛中心距离）距离归一化的结果进行比较，我们在表格1上报告了两个归一化因子的结果。 实施细节。裁剪所有训练图像
表1：300 W公共子集，具有挑战性的子集和全集（68个标志）的平均误差（％）
表2：300-W测试集（68个标志）的平均误差（％） 准确度报告为AUC和失败率。并根据提供的边界框调整为256times;256。如果在我们的实验中没有特别指出，估算器会堆叠四次。对于消融研究，由于考虑时间和计算成本，估计器被堆叠两次。我们所有的模型都经过Caffe培训[24] 在4个Titan X GPU上。请注意，根据提供的边界框裁剪和调整所有测试图像的大小，而不进行任何空间转换，以便与其他方法进行公平比较。对于有限的纸张 剩余内容已隐藏，支付完成后下载完整资料 资料编号：[609947]，资料为PDF文档或Word文档，PDF文档可免费转换为Word </p |
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
以上是毕业论文外文翻译，课题毕业论文、任务书、文献综述、开题报告、程序设计、图纸设计等资料可联系客服协助查找。
