AdaptiveWingLoss

arXiv

Pytorch Implementation of Adaptive Wing Loss for Robust Face Alignment via Heatmap Regression.

Update Logs:

October 28, 2019

• Pretrained Model and evaluation code on WFLW dataset is released.

Installation

Note: Code was originally developed under Python2.X and Pytorch 0.4. This released version was revisioned from original code and was tested on Python 3.7 and Pytorch 1.4.0.

Linux

Install system requirements:

sudo apt-get install python3-dev python3-pip python3-tk libglib2.0-0

Windows

It is recommended use Anaconda.

conda create -n face-lands python=3.7 pip
conda activate face-lands

Install PyTorch 1.4.0

Install python dependencies:

pip install -r requirements.txt

Predict landmarks

1. Download pretrained model from Google Drive and put it in ./ckpt directory.
2. Copy your dataset with faces at ./Dataset directory. Structure of directories can be any. All *.png, *.jpg will be loaded. You can set own extensions by setting --img-extensions option.
3. Run command:

python ./eval.py --val_img_dir ./Dataset  --pred <path_to_file>/pred_landmarks.json' --hg_blocks=4 --pretrained_weights=./ckpt/WFLW_4HG.pth'

Notes

Pretrained model predicts 98 landmarks. pred_landmarks.json is json with relative image paths to val_img_dir. Coordinates of landmarks is not scaled. Predicted coordinates of landmarks placed on the image 64x64.

All input images resized to 256x256.

You can save images with landmarks if set --debug-dir option. A path to directory.

--gray-scale convert all input images to grayscale.

--detect-face auto crop face on loaded image, but recommended to make preprocessing and does not set this option.

Additional options: --help.

Run Evaluation on WFLW dataset

1. Download and process WFLW dataset

   • Download WFLW dataset and annotation from Here.
   • Unzip WFLW dataset and annotations and move files into ./dataset directory. Your directory should look like this:

     AdaptiveWingLoss
     └───dataset
        │
        └───WFLW_annotations
        │   └───list_98pt_rect_attr_train_test
        │   │
        │   └───list_98pt_test
        │
        └───WFLW_images
            └───0--Parade
            │
            └───...
     

   • Inside ./dataset directory, run:

     python convert_WFLW.py
     

     A new directory ./dataset/WFLW_test should be generated with 2500 processed testing images and corresponding landmarks.

2. Download pretrained model from Google Drive and put it in ./ckpt directory.

3. Within ./Scripts directory, run following command:

   sh eval_wflw.sh
   

   *GTBbox indicates the ground truth landmarks are used as bounding box to crop faces.

Future Plans

• Release evaluation code and pretrained model on WFLW dataset.

• Release training code on WFLW dataset.

• Release pretrained model and code on 300W, AFLW and COFW dataset.

• Replease facial landmark detection API

Citation

If you find this useful for your research, please cite the following paper.

@InProceedings{Wang_2019_ICCV,
author = {Wang, Xinyao and Bo, Liefeng and Fuxin, Li},
title = {Adaptive Wing Loss for Robust Face Alignment via Heatmap Regression},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {October},
year = {2019}
}

Acknowledgments

This repository borrows or partially modifies hourglass model and data processing code from face alignment and pose-hg-train.