Skip to content

Latest commit

 

History

History
129 lines (106 loc) · 5.14 KB

README.md

File metadata and controls

129 lines (106 loc) · 5.14 KB

CPM for facial landmark detection

Introduction

This is a pytorch version of Convolutional Pose Machines for facial landmarks detection,you can find more details in original paper for Convolutional Pose Machines

Results

abc

Requirements

  • Python3.5
  • Pytorch 0.4.0
  • torchvision

Dataset Preparation

The raw dataset should be put into the datasets folder.

Dataset Format

Each dataset is saved as one file, in which each row indicates one specific face in one image or one video frame. The format of one line :

image_path annotation_path x1 y1 x2 y2 (face_size)
  • image_path: the image file path of that face.
  • annotation_path: the annotation file path of that face (annotation is the coordinates of all landmarks)
  • x1, y1, x2, y2: the coordinates of left-upper and right-lower points of the face bounding box.
  • face_size: an optional item. If set this value, we use the face_size to compute the NME; otherwise, we use the distance between two pre-defined points to compute the NME.

300-W

Download

unzip ibug.zip -d ibug
mv ibug/image_092\ _01.jpg ibug/image_092_01.jpg
mv ibug/image_092\ _01.pts ibug/image_092_01.pts

unzip afw.zip -d afw
unzip helen.zip -d helen
unzip lfpw.zip -d lfpw
unzip bounding_boxes.zip ; mv Bounding\ Boxes Bounding_Boxes

The 300W directory is in ./datasets/300W and the sturecture is:

.
├── afw
├── afw.zip
├── Bounding_Boxes
├── bounding_boxes.zip
├── helen
├── helen.zip
├── ibug
├── ibug.zip
├── lfpw
└── lfpw.zip

Then you use the script to generate the 300-W list files.

python generate_300W.py

All list files will be saved into ./datasets/300W_lists. The files *.DET use the face detecter results for face bounding box. *.GTB use the ground-truth results for face bounding box.

can not find the *.mat files for 300-W.

The download link is in the official 300-W website.

https://ibug.doc.ic.ac.uk/media/uploads/competitions/bounding_boxes.zip

The zip file should be unzipped, and all extracted mat files should be put into ./datasets//300W/Bounding_Boxes.

AFLW

download AFLW datasets from the official website,you can alse download it from my baiduyun aflw-images-0.tar.gz --Baiduyun Password: oi38 aflw-images-2.tar.gz --Baiduyun Password: u39h aflw-images-3.tar.gz --Baiduyun Password: ka2b

Download the AFLW datasets in datasets/AFLW and extract it by tar xzvf aflw-images-*.tar.gz.

mv aflw/data/flickr/0 images/
mv aflw/da1ta/flickr/2 images/
mv aflw/data/flickr/3 images/

Download the AFLWinfo_release.mat from this website into ./cache_data. This is the revised annotation of the full AFLW dataset. put AFLWinfo_release.mat file in datasets/AFLW folder, and he structure of AFLW is:

.
├── AFLWinfo_release.mat
└── images
    ├── 0
    ├── 2
    └── 3

use follow script to generate the AFLW dataset list file into ./datasets/AFLW_lists.

python generate_AFLW.py

Train and Eval

set the training parameter in configs/Detector.config ,configs/SGD.config and script/300W-DET.sh

bash script/300W-DET-TRAIN.sh  #for trainning 300W

bash script/AFLW-DET-TRAIN.sh  #for trainning AFLW

ash script/300W-DET-EVAL.sh  #for evaluating 300W

bash script/AFLW-DET-EVAL.sh  #for evaluating AFLW

visualize

Pre-trained model which trained on 300W datasets can be downloaded from here 300w_68pts_cpm_vgg16-epoch-049-050.pth --Baiduyun passwd:fe6e

Pre-trained model which trained on AFLW datasets can be downloaded from here AFLW_19pts_cpm_vgg16-epoch-049-050.pth --Baiduyun passwd:sia1

python ./demo.py --model cpm_vgg16-epoch-049-050.pth --image datasets/images/image_0019.png

Note

the repository is based on this repo