已下载:
一、300W
http://www.cbsr.ia.ac.cn/users/xiangyuzhu/projects/3DDFA/main.htm
三维密集人脸对齐(3D Dense Face alignment, 3DDFA),将稠密的三维人脸模型拟合到图像中。
[Face Profiling]: The MATLAB code to synthesize large-pose face appearances with 3D Face Model.
[3DDFA]: The MATLAB code to fit 3DMM to a face image across large poses.
[300W-3D]: The fitted 3D Morphable Model (3DMM) parameters of 300W samples.
[300W-3D-Face]:The fitted 3D mesh, which is needed if you do not have Basel Face Model (BFM)
[300W-LP]: The synthesized large-pose face images from 300W.
[AFLW2000-3D]:The fitted 3D faces of the first 2000 AFLW samples, which can be used for 3D face alignment evaluation.
http://www.cbsr.ia.ac.cn/users/xiangyuzhu/projects/3ddfa/Database/AFLW2000-3D.zip
二、Spacetime Faces: Spacetime Faces
《Spacetime Faces: High-Resolution Capture for Modeling and Animation》
384个face mesh的序列,ply格式,每个mesh 23728个顶点。
mesh的索引从306到1455,索引增加3。
http://grail.cs.washington.edu/projects/stfaces/data/face205_ply.zip
三、 LS3D-W(人脸图像来自AFLW, 300VW, 300W和FDDB)
只提供3D landmark
The balanced subset https://www.adrianbulat.com/downloads/FaceAlignment/LS3D-W-balanced-20-03-2017.zip
demo code GitHub - 1adrianb/2D-and-3D- face-alignment
需申请:
一、 FRGC2
FRVT 2006 and ICE 2006 large-scale experimental results (depth and texture image) 需申请
二、BU-3DFE *** BFM2017用到
A 3D facial expression database for facial behavior research (mesh and texture image) 需申请
3D Facial Expression Database - Binghamton University
三、 Bosphorus
Bosphorus database for 3D face analysis (depth and texture) 需申请
四、Florence
The Florence 2D/3D Hybrid Face Dataset (triangle mesh and texture) 需申请
MICC - Media Integration and Communication Center
五、FaceWarehouse
FaceWarehouse: A 3D facial expression database for visual computing (640*480 RGBD + triangle mesh )需申请
六 、CASIA 3D Face Database
(640×480 depth map and texture image)
七、Biwi 3D Audiovisual Corpus of Affective Communication (B3D(AC)ˆ2)
(triangle mesh and 780×580 texture image +1024×768 UV texture map) 需申请
八、D3DFACS - Dynamic 3D Facial Action Coding System (triangle mesh and 1024×1280 UV texture map)
九、 Headspace (triangle mesh, 2973×3055 UV texture map) 需申请
A 3D Morphable Model of Craniofacial Shape and Texture Variation
十、CoMA (triangle mesh, texture images) 需申请
Generating 3D faces using Convolutional Mesh Autoencoders
十一、VOCASET (triangle mesh, texture images) 需申请
Capture, Learning, and Synthesis of 3D Speaking Styles
十二、4DFAB
4DFAB: A Large Scale 4D Database for Facial Expression Analysis and Biometric Applications (triangle mesh and UV texture)
十三、York 3D Face Database (depth map and texture image) 需申请
Three-dimensional face recognition using combinations of surface feature map subspace components
十四、3DRFE (triangle mesh + 1296×1944 diffuse and specular albedo maps + hybrid normal maps)
Effect of illumination on automatic expression recognition: a novel 3D relightable facial database
十五、BP4D-Spontaneous (triangle mesh + 1040 × 1,329 texture image) 需申请
BP4D-Spontaneous: a high-resolution spontaneous 3D dynamic facial expression database
3D Facial Expression Database - Binghamton University
2D image生成3D groundtruth
《Disentangling Features in 3D Face Shapes for Joint Face Reconstruction and Recognition》从二维人脸识别数据库CASIA-Webface构建训练数据,多张图片拟合3DMM based on BFM。 MICC(3D mesh),BU-3DFE(mesh),LFW(Labeled face in the wild)和YTF(Youtub Face)数据库全面实验。
《Regressing robust and discriminative 3D morphable models with a very deep neural network》参考Automated 3D Face Reconstruction from Multiple Images using Quality Measures。在CASIA-Webface数据库,先对数据集中500k张图像进行3DMM重建,得到各自的相关参数(主要是形状和纹理,其次还有相机参数和光照);再对属于同一个 object 进行处理,使得同一个 object 的三维结果一致,一般是取平均。
《Self-Supervised Adaptation of High-Fidelity Face Models for Monocular Performance Tracking》 收集新的数据集,高分辨率(1920×1080)、in the wild、3D mesh,base on 受控实验室环境中捕获的高质量人脸模型《Deep appearance models for face rendering》
虚拟人脸生成
***1、3D face reconstruction by learning from synthetic data
弱透视投影和Phong反射模型来渲染合成图像。
***2、SfSNet: Learning shape, re ectance and illuminance of faces “in the wild”
SfSNet通过输入一张人脸图像,从中恢复形状,反射和光照这三方面的信息
使用虚拟数据,3D建模的各种脸,记录了它们的Ground-Truth normal, albedo 和 lighting,
同时使用真实的人脸图像CelebA dataset,有RGB图像,没有GT的形状,反射,和光照信息。
在有标签的合成数据上训练一个简单的基于跳跃连接的编码—解码网络。将此网络应用于真实数据,以获得法向,反照率和照明估计,在下一阶段用作“伪监督”。
用带有GT标签的合成数据和带有“伪监督”标签的真实数据训练SfSNet。 损失除了法向,反照率和光照的监督损失,还有重建损失,其旨在最小化原始图像和重建图像之间的误差。、
重构模型与合成图像结合
1、Face normals “in-the-wild" using fully convolutional networks
使用LSFM作为3DMM,并使用FaceWarehouse模型添加表情。基于ICT-3DRFE数据库和Photoface数据库构建训练数据集,以生成改变光照的合成图像。用同样的模型,他们随机生成合成的3D人脸,这些人脸与野外图像对齐,为渲染提供真实的背景。
ICT-3DRFE: effiect of illumination on automatic expression recognition: a novel 3D relightable facial database
Photoface: The Photoface database
2、Deep monocular inverse face rendering
使用虚拟人脸构建了第一个训练集,用于对网络进行预训练。然后,在迭代训练期间,使用预训练的网络从真实图像中估计模型参数。在绘制合成图像之前,将高斯噪声添加到这些估计参数中,用于在下一次迭代中训练网络。
自带数据集的论文
1、Dual neural networks coupling data regression with explicit priors for monocular 3D face reconstruction(未找到)
用3DMM拟合二维图像,再与三角化深度图像进行非刚性配准,获得相同三角化的真值三维人脸
2 、Beyond 3DMM space: Towards fine-grained 3D face reconstruction
对RGB-D图像进行3DMM拟合,根据几何和纹理的对应关系进行非刚性配准。
3、Multi-view 3D face reconstruction in the wild using siamese networks (未找到)
6000多人捕捉2D图像和3D人脸
4、AvatarMe: Realistically renderable 3D facial reconstruction "in-the-wild"
200多人在7种不同表情下的3D脸和reflectance maps
RealFaceDB 待开源
5、Lightweight photometric stereo for facial details recovery
84名不同种族、性别和年龄的受试者,每个受试者有29种不同的表情。1600*1200分辨率。利用在不同近点光照位置和光照条件下拍摄的图像重建具有丰富细节的三维人脸模型。
数据增强:
(a)从3DMM中随机生成形状和表达式参数生成不同的图形,并将在真实数据集中获得的反照率进行转换。
(b)使用非刚性ICP将Light Stage 数据集中的人脸模型与平均形状进行拟合,并结合真实数据集中的反照率生成训练数据。
Light Stage : Rapid acquisition of specular and diffuse normal maps from polarized spherical gradient illumination
一个公开的数据集,包含了23个人,他们拥有15种不同的表情和相应的高分辨率3D模型(未找到)
3D human body dataset