- Hugging Face: download link
assets
├── face_model.npy # (face model)
├── large_base_net.pth
├── net_recon.pth # (backbone)
├── net_recon_mbnet.pth # (MobileNet-V3 backbone, optional)
├── retinaface_resnet50_2020-07-20_old_torch.pth
├── similarity_Lm3D_all.mat
├── indices_38365_35709.npy # (optional)
├── indices_53215_35709.npy # (optional)
├── indices_53215_38365.npy # (optional)
├── indices_53490_35709.npy # (optional)
├── meanshape-68ldms.obj # (optional)
├── meanshape-106ldms.obj # (optional)
├── meanshape-134ldms.obj # (optional)
├── meanshape-parallel.obj # (optional)
└── meanshape-seg.obj # (optional)
-
Load the masks (annotations) using the following method:
import numpy as np model = np.load("./assets/face_model.npy",allow_pickle=True).item()
model['annotation']
andmodel['annotation_tri']
: segmentation annotation indices and triangle faces for 8 parts.model['ldm106']
: vertex indices for 106 landmarks.model['parallel']
: parallel for 33 face profile landmarks, used for dynamic 2D and 3D landmark marchingmodel['ldm134']
: vertex indices for 134 landmarks.model['ldm68']
: vertex indices for 68 landmarks. -
The commonly used BFM has 4 different topological structures with 53,490 / 53,215 / 38,365 / 35,709 vertices. We summarize the correspondence indices and provide them in the form of
indices_A_B.npy
, whereA
andB
represent the vertex numbers (index starts from 0). These indices and obj files are not essential for running 3DDFA_V3, but they may help beginners better and more quickly use BFM in some cases.
net_recon.pth
is the checkpoint of our 3DDFA_V3.face_model.npy
is the face model and attributes based on 3DDFA_V3, BFM, Exp_Pca, and Deep3D.large_base_net.pth
andretinaface_resnet50_2020-07-20_old_torch.pth
are used for face detector from HRN and retinaface.similarity_Lm3D_all.mat
is used for cropping faces from Deep3D.