face-landmarks-dlib

Face landmarks detection with DLib

Data preparation

To download 300W and Menpo datasets simply run:

bash scripts/download_data.sh

Then prepare splits for 300W and Menpo datasets in .xml format by running:

bash scripts/prepare_data.sh

or download it:

bash scripts/download_splits.sh

The dataset directory structure should be the following:

data/landmarks_task
├── 300W/  # annotation json files
    ├── train/    # train images
    ├── test/    # test images
    ├── train_with_face_landmarks.xml/    # 300W train annotations
    └── test_with_face_landmarks.xml/    # 300W test annotations
├── Menpo/
    ├── train/    # train images
    ├── test/    # test images
    ├── train_with_face_landmarks.xml/    # Menpo train annotations
    └── test_with_face_landmarks.xml/    # Menpo test annotations
├── train_with_face_landmarks.xml/    # 300W + Menpo train annotations
└── test_with_face_landmarks.xml/    # 300W + Menpo test annotations

Training

To train face shape predictor on 300W + Menpo dataset simply run:

bash scripts/shape_predictor_68_300W_Menpo.sh

python train.py \
    --dataset 300W Menpo \
    --data_root data/landmarks_task \
    --train_annotation train_with_face_landmarks.xml \
    --test_annotation test_with_face_landmarks.xml \
    --save_dir output

Evaluation

Download trained models by running:

bash scripts/download_models.sh

or use your own trained model.

To evaluate model on 300W + Menpo dataset and save predictions run:

bash scripts/shape_predictor_68_300W_Menpo_eval.sh

python evaluate.py \
    --model_path output/shape_predictor_68_300W_Menpo.dat \
    --data_root data/landmarks_task \
    --image_dir test \
    --save_dir results/300W_Menpo

Plot CED curve

Download results of shape_predictor_68_300W_Menpo.dat model on 300W, Menpo and 300W + Menpo test splits:

bash scripts/download_results.sh

or use your own results.

To plot CED curve simply run:

python compute_ced.py \
    --gt_path data/landmarks_task/300W/test data/landmarks_task/Menpo/test \
    --predictions_path results/300W results/Menpo results/300W_Menpo \
    --output_path results/ced.png

Inference

Download trained models by running:

bash scripts/download_models.sh

or use your own trained model.

To visualize results on example images run:

python inference.py \
    --model_path output/shape_predictor_68_300W_Menpo.dat \
    --data_root examples \
    --save_dir examples/300W_Menpo

akanametov / face-landmarks-dlib Goto Github PK