This is the official repository for the following paper:

Generalizable Face Landmarking Guided by Conditional Face Warping [paper] [arxiv] [project page]

Jiayi Liang*, Haotian Liu*, Hongteng Xu, Dixin Luo
Accepted by CVPR 2024.

pip install -r requirements.txt

Our proposed framework mainly contains two parts: face warper and landmark detector. They are trained in an alternative optimization framework.

• The face warper aims to deform real human faces according to stylized facial images, generating warped faces and corresponding warping fields.
• The face landmarker, as the key module of the warper, predicts the ending points of the warping fields and thus provides us with pseudo landmarks for the stylized facial images.

In our implmentation, we employ SLPT as our backbone and locate the model in Landmark2 folder. For the reproduction on other detectors, substitute the Landmark2 folder with target model and make modifications in train.py.

Download images and annotations of 300-W from ibug.

We select frontal faces from the trainset of 300W as our training data, and list of image path 300W_frontal_train_list.txt can be downloaded from Google Drive.

Your directory should be like:

  Dataset
  │
  └──300W
     │
     └───300W_frontal_train_list.txt
     └───frontal_train
         └───261068_1.jpg
         │
         └───...
     └───frontal_train_label
         └───261068_1.jpg.npy
         │
         └───...
     └───test_list.txt
     └───test_list_common.txt
     └───test_list_challenge.txt
     └───lfpw
         └───trainset
         └───testset
             └───image_0001.png
             └───image_0001.pts
             │
             └───...
     └───helen
         │
         └───...
     └───ibug
         │
         └───...

• Download CariFace according to CariFace Dataset. The split of training and testing set can be downloaded from Google Drive.

• Other domains like Artistic-Faces can be retrieved from Link. The split of training and testing set can be downloaded from Google Drive.

Please note that ArtiFace contains a total of 160 images. Under the GZSL (Unseen ArtiFace) scenario, the test set size of ArtiFace is 160 images; under the DA (300W->ArtiFace) and GZSL (Unseen CariFace) scenarios, the test set of ArtiFace only contains 32 images.
Therefore, please modify the specified test set list in Artistic.py to test_list.txt (32 images) or test_list_all.txt (160 images) according to the different circumstances.

Your directory should be like:

  Dataset
  │
  └──CariFace_dataset
     │
     └───images
         └───00005.jpg
         │
         └───...
     └───landmarks
         └───00005.jpg.npy
         │
         └───...
     └───train_list.txt
     └───test_list.txt
  │
  └──AF_dataset
     │
     └───images
         └───0.png
         │
         └───...
     └───landmarks
         └───0.png.npy
         │
         └───...
     └───train_list.txt
     └───test_list.txt
     └───test_list_all.txt

Download source-pretrained weights model_best.pth from Google Drive and move it into folder Landmark2.

python train.py --src_data path/to/source/data --tgt_data path/to/target/data --pretrain_path path/to/pretrained/checkpoint

Download our model and test on the CariFace by running:

python test.py --checkpoint path/to/model/weights

Further, to test on ArtiFace, download checkpoint and inference:

python test_Artistic.py --checkpoint path/to/model/weights

If our work is helpful for your research, please cite our paper: