Model
1、PRNet
http://openaccess.thecvf.com/content_ECCV_2018/papers/Yao_Feng_Joint_3D_Face_ECCV_2018_paper.pdf
https://github.com/YadiraF/PRNet
image.png
https://www.jianshu.com/p/c5e6820a599a
PRNet如和获取特征点：
模型推导出3D position map后选取指定位置的点作为特征点
代码在 api.py中
def __init__():
...
self.uv_kpt_ind = np.loadtxt(prefix + '/Data/uv-data/uv_kpt_ind.txt').astype(np.int32) # 2 x 68 get kpt
...
def get_landmarks(self, pos):
'''
Args:
pos: the 3D position map. shape = (256, 256, 3).
Returns:
kpt: 68 3D landmarks. shape = (68, 3).
'''
kpt = pos[self.uv_kpt_ind[1,:], self.uv_kpt_ind[0,:], :]
return kpt
uv_kpt_ind.txt
1.500000000000000000e+01 2.200000000000000000e+01 2.600000000000000000e+01 3.200000000000000000e+01 4.500000000000000000e+01 6.700000000000000000e+01 9.100000000000000000e+01 1.120000000000000000e+02 1.280000000000000000e+02 1.430000000000000000e+02 1.640000000000000000e+02 1.880000000000000000e+02 2.100000000000000000e+02 2.230000000000000000e+02 2.290000000000000000e+02 2.330000000000000000e+02 2.400000000000000000e+02 5.800000000000000000e+01 7.100000000000000000e+01 8.500000000000000000e+01 9.700000000000000000e+01 1.060000000000000000e+02 1.490000000000000000e+02 1.580000000000000000e+02 1.700000000000000000e+02 1.840000000000000000e+02 1.970000000000000000e+02 1.280000000000000000e+02 1.280000000000000000e+02 1.280000000000000000e+02 1.280000000000000000e+02 1.170000000000000000e+02 1.220000000000000000e+02 1.280000000000000000e+02 1.330000000000000000e+02 1.380000000000000000e+02 7.800000000000000000e+01 8.600000000000000000e+01 9.500000000000000000e+01 1.020000000000000000e+02 9.600000000000000000e+01 8.700000000000000000e+01 1.530000000000000000e+02 1.600000000000000000e+02 1.690000000000000000e+02 1.770000000000000000e+02 1.680000000000000000e+02 1.590000000000000000e+02 1.080000000000000000e+02 1.160000000000000000e+02 1.240000000000000000e+02 1.280000000000000000e+02 1.310000000000000000e+02 1.390000000000000000e+02 1.460000000000000000e+02 1.370000000000000000e+02 1.320000000000000000e+02 1.280000000000000000e+02 1.230000000000000000e+02 1.180000000000000000e+02 1.100000000000000000e+02 1.220000000000000000e+02 1.280000000000000000e+02 1.330000000000000000e+02 1.450000000000000000e+02 1.320000000000000000e+02 1.280000000000000000e+02 1.230000000000000000e+02
9.600000000000000000e+01 1.180000000000000000e+02 1.410000000000000000e+02 1.650000000000000000e+02 1.830000000000000000e+02 1.900000000000000000e+02 1.880000000000000000e+02 1.870000000000000000e+02 1.930000000000000000e+02 1.870000000000000000e+02 1.880000000000000000e+02 1.900000000000000000e+02 1.830000000000000000e+02 1.650000000000000000e+02 1.410000000000000000e+02 1.180000000000000000e+02 9.600000000000000000e+01 4.900000000000000000e+01 4.200000000000000000e+01 3.900000000000000000e+01 4.000000000000000000e+01 4.200000000000000000e+01 4.200000000000000000e+01 4.000000000000000000e+01 3.900000000000000000e+01 4.200000000000000000e+01 4.900000000000000000e+01 5.900000000000000000e+01 7.300000000000000000e+01 8.600000000000000000e+01 9.600000000000000000e+01 1.110000000000000000e+02 1.130000000000000000e+02 1.150000000000000000e+02 1.130000000000000000e+02 1.110000000000000000e+02 6.700000000000000000e+01 6.000000000000000000e+01 6.100000000000000000e+01 6.500000000000000000e+01 6.800000000000000000e+01 6.900000000000000000e+01 6.500000000000000000e+01 6.100000000000000000e+01 6.000000000000000000e+01 6.700000000000000000e+01 6.900000000000000000e+01 6.800000000000000000e+01 1.420000000000000000e+02 1.310000000000000000e+02 1.270000000000000000e+02 1.280000000000000000e+02 1.270000000000000000e+02 1.310000000000000000e+02 1.420000000000000000e+02 1.480000000000000000e+02 1.500000000000000000e+02 1.500000000000000000e+02 1.500000000000000000e+02 1.480000000000000000e+02 1.410000000000000000e+02 1.350000000000000000e+02 1.340000000000000000e+02 1.350000000000000000e+02 1.420000000000000000e+02 1.430000000000000000e+02 1.420000000000000000e+02 1.430000000000000000e+02
获取的特征点为3维的，所以可以通过坐标系旋转对人脸和特征点进行旋转，从而获得旋转后的人脸和特征点数据。
效果如下：左边是原图，右边为生成的3D模型
image.png
image.png
感觉不太像，可能和训练数据集有关系。虽然不太像，但毕竟是人脸，应该可以用来构造多角度数据。
2、In The Wild 3D Morphable Models
3D Reconstruction of "In-the-Wild" Faces in Images and Videos --- J. Booth, A. Roussos, E. Ververas, E. Antonakos, S. Ploumpis, Y. Panagagakis, S. Zafeiriou. Transactions on Pattern Analysis and Machine Intelligence (T-PAMI), accepted for publication (2018).
https://github.com/menpo/itwmm
3、https://github.com/YadiraF/face3d
Python tools for 3D face: 3DMM, Mesh processing(transform, camera, light, render), 3D face representations.
Dataset
https://ibug.doc.ic.ac.uk/resources/300-VW/