（简单易懂PRNet）Joint 3D Face Reconstruction and Dense Alignment with position Map Regression Netw

野性酷女 2022-09-01 12:59 192阅读 0赞

原文链接：\[https://blog.csdn.net/yfy1127yfy/article/details/102839840?ops\_request\_misc=%257B%2522request%255Fid%2522%253A%2522162736575416780261967321%2522%252C%2522scm%2522%253A%252220140713.130102334.pc%255Fall.%2522%257D&request\_id=162736575416780261967321&biz\_id=0&utm\_medium=distribute.pc\_search\_result.none-task-blog-2allfirst\_rank\_v2~rank\_v29-6-102839840.first\_rank\_v2\_pc\_rank\_v29&utm\_term=Joint+3D+Face+Reconstruction+and+Dense+Alignment+with+Position+Map+Regression+Network&spm=1018.2226.3001.4187\]  
(https://blog.csdn.net/yfy1127yfy/article/details/102839840?ops\_request\_misc=%257B%2522request%255Fid%2522%253A%2522162736575416780261967321%2522%252C%2522scm%2522%253A%252220140713.130102334.pc%255Fall.%2522%257D&request\_id=162736575416780261967321&biz\_id=0&utm\_medium=distribute.pc\_search\_result.none-task-blog-2allfirst\_rank\_v2~rank\_v29-6-102839840.first\_rank\_v2\_pc\_rank\_v29&utm\_term=Joint%203D%20Face%20Reconstruction%20and%20Dense%20Alignment%20with%20Position%20Map%20Regression%20Network&spm=1018.2226.3001.4187)  
**Github**项目链接：https://github.com/YadiraF/PRNet  
**训练数据300W-LP，ground-truth是没有用UV表示的，所以要用3DMM系数先生成UV训练数据。再根据损失函数，最后模型输出的也是UV-map图。**

2019/10/31,想起来之前看的没做笔记，重新总结下吧。

论文地址：[http://openaccess.thecvf.com/content\_ECCV\_2018/papers/Yao\_Feng\_Joint\_3D\_Face\_ECCV\_2018\_paper.pdf][http_openaccess.thecvf.com_content_ECCV_2018_papers_Yao_Feng_Joint_3D_Face_ECCV_2018_paper.pdf]

github地址：[https://github.com/YadiraF/PRNet][https_github.com_YadiraF_PRNet]

# 摘要 #

本文提出一种直接的方法，来实现3D人脸重建和稠密对齐。设计了一种2D表达，称之为UV坐标图，可以在UV空间保存完整的人脸3D形状。然后就可以训练CNN网络从一张2D图回归出3D形状。

# 一、简介 #

早期的3D重建和稠密对齐研究中，使用过2D关键点，但2D关键点无法处理大角度和遮挡；

使用了3DMM方法，又受限于透视投影和3D样条曲线（3D ThinPlate Spline）的计算量大；

有的端到端方案，摆脱了这些限制，但需要额外的网络来估计深度信息，且并没有提供稠密对齐；

VRN提出了体素表示法，该方法计算量大，分辨率低，且由于点云稀疏特性，存在大量无效计算。

本文提出了一种端到端的PRN多任务方法，能够同时完成稠密人脸对齐和3D人脸形状重建。主要几个贡献：

1.  以端到端方式，实现了高分辨率的3D人脸重建和稠密对齐；
2.  设计了UV位置图，来记录人脸的3D位置信息；
3.  设计了权重掩模用于loss计算，loss中每个点权重不同，可明显提高网络性能
4.  CNN采用轻量级模式，单张人脸任务可达到100FPS
5.  在AFLW200-3D和Florence数据集上可达到25%的性能提升

# 二、方法介绍 #

拿之前写的ppt用一下吧。

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lmeTExMjd5Znk_size_16_color_FFFFFF_t_70][] （a）输入图+点云；（b）输入图；（c）UV纹理图；（d）UV坐标图；（e）（f）（g）UV坐标的三通道  
图1 UV空间

开源数据集，如300W-LP，ground-truth是没有用UV表示的，所以要先生成UV训练数据。将（a）中的点云坐标转变成（d）的表达形式，该方法如图2：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lmeTExMjd5Znk_size_16_color_FFFFFF_t_70 1][] 图2 点云转换为UV表示

生成需要的训练数据后就可以用轻量级的CNN网络来处理了：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lmeTExMjd5Znk_size_16_color_FFFFFF_t_70 2][] 图3 PRN网络结构

网络结构代码如下：

se = tcl.conv2d(x, num_outputs=size, kernel_size= 4, stride= 1) # 256 x 256 x 16 
         se = resBlock(se, num_outputs=size * 2, kernel_size= 4, stride= 2) # 128 x 128 x 32 
         se = resBlock(se, num_outputs=size * 2, kernel_size= 4, stride= 1) # 128 x 128 x 32 
         se = resBlock(se, num_outputs=size * 4, kernel_size= 4, stride= 2) # 64 x 64 x 64 
         se = resBlock(se, num_outputs=size * 4, kernel_size= 4, stride= 1) # 64 x 64 x 64 
         se = resBlock(se, num_outputs=size * 8, kernel_size= 4, stride= 2) # 32 x 32 x 128 
         se = resBlock(se, num_outputs=size * 8, kernel_size= 4, stride= 1) # 32 x 32 x 128 
         se = resBlock(se, num_outputs=size * 16, kernel_size= 4, stride= 2) # 16 x 16 x 256 
         se = resBlock(se, num_outputs=size * 16, kernel_size= 4, stride= 1) # 16 x 16 x 256 
         se = resBlock(se, num_outputs=size * 32, kernel_size= 4, stride= 2) # 8 x 8 x 512 
         se = resBlock(se, num_outputs=size * 32, kernel_size= 4, stride= 1) # 8 x 8 x 512 
         
         pd = tcl.conv2d_transpose(se, size * 32, 4, stride= 1) # 8 x 8 x 512 
         pd = tcl.conv2d_transpose(pd, size * 16, 4, stride= 2) # 16 x 16 x 256 
         pd = tcl.conv2d_transpose(pd, size * 16, 4, stride= 1) # 16 x 16 x 256 
         pd = tcl.conv2d_transpose(pd, size * 16, 4, stride= 1) # 16 x 16 x 256 
         pd = tcl.conv2d_transpose(pd, size * 8, 4, stride= 2) # 32 x 32 x 128 
         pd = tcl.conv2d_transpose(pd, size * 8, 4, stride= 1) # 32 x 32 x 128 
         pd = tcl.conv2d_transpose(pd, size * 8, 4, stride= 1) # 32 x 32 x 128 
         pd = tcl.conv2d_transpose(pd, size * 4, 4, stride= 2) # 64 x 64 x 64 
         pd = tcl.conv2d_transpose(pd, size * 4, 4, stride= 1) # 64 x 64 x 64 
         pd = tcl.conv2d_transpose(pd, size * 4, 4, stride= 1) # 64 x 64 x 64 
         pd = tcl.conv2d_transpose(pd, size * 2, 4, stride= 2) # 128 x 128 x 32 
         pd = tcl.conv2d_transpose(pd, size * 2, 4, stride= 1) # 128 x 128 x 32 
         pd = tcl.conv2d_transpose(pd, size, 4, stride= 2) # 256 x 256 x 16 
         pd = tcl.conv2d_transpose(pd, size, 4, stride= 1) # 256 x 256 x 16 
         pd = tcl.conv2d_transpose(pd, 3, 4, stride= 1) # 256 x 256 x 3 
         pd = tcl.conv2d_transpose(pd, 3, 4, stride= 1) # 256 x 256 x 3 
         pos = tcl.conv2d_transpose(pd, 3, 4, stride= 1, activation_fn = tf.nn.sigmoid)

残差块代码如下，激活是relu，归一化是BN，shortcut对应三次卷积，随后通道合并，最后归一化和激活：

def resBlock(x, num_outputs, kernel_size = 4, stride=1, activation_fn=tf.nn.relu, normalizer_fn=tcl.batch_norm, scope=None): 
         assert num_outputs% 2== 0 #num_outputs must be divided by channel_factor(2 here) 
         with tf.variable_scope(scope, 'resBlock'): 
         shortcut = x 
         if stride != 1 or x.get_shape()[ 3] != num_outputs: 
         shortcut = tcl.conv2d(shortcut, num_outputs, kernel_size= 1, stride=stride, 
         activation_fn= None, normalizer_fn= None, scope= 'shortcut') 
         x = tcl.conv2d(x, num_outputs/ 2, kernel_size= 1, stride= 1, padding= 'SAME') 
         x = tcl.conv2d(x, num_outputs/ 2, kernel_size=kernel_size, stride=stride, padding= 'SAME') 
         x = tcl.conv2d(x, num_outputs, kernel_size= 1, stride= 1, activation_fn= None, padding= 'SAME', normalizer_fn= None) 
         
         x += shortcut 
         x = normalizer_fn(x) 
         x = activation_fn(x) 
         return x

## Loss ##

![20191031173238166.png][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lmeTExMjd5Znk_size_16_color_FFFFFF_t_70 3][]

P(x,y) 是UV位置空间的预测结果，代表着每个像素的xyz位置  
W(x,y) 是UV位置空间的权重，对UV空间进行权重控制，2D关键点:眼鼻嘴:脸部其他:其他 = 16:4:3:0；

## 训练 ##

训练数据来源: 使用了300W-LP数据集

*  拥有各个角度的人脸数据，resize到256x256
 *  3DMM系数的标注
 *  使用3DMM 生成3D点云，并转换3D点云至UV空间

虽然生成GT使用了3DMM的标注系数,但是模型本身不包含3DMM模型的任何线性约束.

数据增广：各种场景

*  角度变换：-45 ~ 45 度
 *  平移：系数0.9 ~ 1.2 (原图大小为基准)
 *  颜色通道变换：系数0.6 ~ 1.4
 *  添加噪音、纹理遮挡,模拟真实情况遮挡.
 *  adam优化器,初始学习率0.0001,每5个epoch,衰减1半,batch size:16

# 三、测试结果 #

先提一下这个方法能做的事，由于学习到了2D与3D之间的映射，因次能实现的功能如下：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lmeTExMjd5Znk_size_16_color_FFFFFF_t_70 4][]

论文中部分测试指标如下：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lmeTExMjd5Znk_size_16_color_FFFFFF_t_70 5][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lmeTExMjd5Znk_size_16_color_FFFFFF_t_70 6][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lmeTExMjd5Znk_size_16_color_FFFFFF_t_70 7][]

[http_openaccess.thecvf.com_content_ECCV_2018_papers_Yao_Feng_Joint_3D_Face_ECCV_2018_paper.pdf]: http://openaccess.thecvf.com/content_ECCV_2018/papers/Yao_Feng_Joint_3D_Face_ECCV_2018_paper.pdf
[https_github.com_YadiraF_PRNet]: https://github.com/YadiraF/PRNet
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lmeTExMjd5Znk_size_16_color_FFFFFF_t_70]: /images/20220829/1c637542fc8b442b8f4a22c6c4456bd2.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lmeTExMjd5Znk_size_16_color_FFFFFF_t_70 1]: /images/20220829/26afbfa9bd004ec6bf5085a692e360e4.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lmeTExMjd5Znk_size_16_color_FFFFFF_t_70 2]: /images/20220829/e5c7f6c8355a44f5b4fc62fdf10112b9.png
[20191031173238166.png]: /images/20220829/b469a56790b14df29eb64657dc1ad1e0.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lmeTExMjd5Znk_size_16_color_FFFFFF_t_70 3]: /images/20220829/03a70e65888f40ef9ee9a90c642b697d.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lmeTExMjd5Znk_size_16_color_FFFFFF_t_70 4]: /images/20220829/1f9e83ef0ad14a00a0bbca0d97057dad.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lmeTExMjd5Znk_size_16_color_FFFFFF_t_70 5]: /images/20220829/70c0c19c7cc54be19785540e0b2b4610.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lmeTExMjd5Znk_size_16_color_FFFFFF_t_70 6]: /images/20220829/97247107ed044b19bcae0423238e6bff.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lmeTExMjd5Znk_size_16_color_FFFFFF_t_70 7]: /images/20220829/701a9a61592f47d0839914223d5f5661.png