Description

This is the official implementation of paper Landmark Detection with Learnable Connectivity Graph Convolutional Network

Dependencies

Install these libraries with anaconda and pip

1. pytorch 1.8.0 cuda10.2
2. OpenCV 4.5.1
3. scikit-learn
4. easydict
5. tqdm
6. python 3.8
7. yacs
8. wandb

Training data is saved using wandb. Follow the instruction to create an account and login. In case you don't want to log your training data to wandb, enable offline training option by using these command in terminal

wandb offline

Pretrained model and datasets

As 300W dataset is the combination of multiple datasets, we provide a single download link for convenience.

Download WFLW dataset from here

Download the pretrained weights and 300W dataset from here

WFLW dataset

Training

python scripts/train_wfw -i [image folder] --annotation [traning annotation file] --test_images [image folder] --test_annotation [test annotation file] --augmentation

Evaluating

python scripts/evaluate_wflw.py -i [image folder] --annotation [test annotation file] --weights [pretrained weights]

Single image prediction

python scripts/visualize_prediction.py -i temp/test/20.png --edge

Use "--edge" to visualize connections between landmarks

300W dataset

Training

python python scripts/train_300w.py --annotation [dataset folder]

Evaluating

python scripts/evaluate_300w.py -i [image folder] --annotation [test annotation file] --weights [pretrained weights]

Acknowledgement

This repository reuse code from:

• HRNet Facial Landmark Detection: https://github.com/HRNet/HRNet-Facial-Landmark-Detection
• Objects as Points: https://github.com/xingyizhou/CenterNet

LICENSE

Will be released under MIT License