It is an offcial source program about ATF reference to PFLD

And except for ATF, we also have supported AWing, Wing loss and half or quarter channel.

This programe contains 3 networks:

1. OCN = PFLD + AVGPool - auxiliary_network
2. OCN_0.5 & OCN_0.25
3. OCNMP

If you are interested in achieve higher accuracy, you can change the backbone depth or other tricks. OCNMP is the same structure as OCN at inference, you can clip other auxiliary branches to transfer other framework.

All experiments are conducted on per RTX 1080ti. At present, we have tested OCN and OCNMP (both with L1 loss) the accuracy on various benchmarks.

And for AFLW, we are not sure our schema of the normlizated factor is right, so we have not put here.

Also, we have tested the different performance in the WING loss and L2 loss with OCN_0.5 and OCN_0.25 on WFLW benchmark.

We have not tested wing loss in OCN, ATF in OCN_0.5 and OCN_0.25. If you are intrested in implementing them, we weclome the result publiced.

1. Install dependencies

pip install -r requirements.txt

1. You need to download images (300W, AFLW, WFLW, COFW) from official websites and then put them into native_dataset folder for each dataset.

The image path in image folder can find the csv file, and your data directory should look like this:

fast_landmark_alignment_network
-- checkpoints
-- experiments
-- lib
-- logs
-- losses
-- models
-- tools
-- native_dataset
   |-- 300w
   |   |-- annotations
   |   |   |-- face_landmarks_300w_test.csv
   |   |   |-- face_landmarks_300w_train.csv
   |   |   |-- face_landmarks_300w_valid.csv
   |   |   |-- face_landmarks_300w_valid_challenge.csv
   |   |   |-- face_landmarks_300w_valid_common.csv
   |   |-- images
   |   |   |-- aflw
   |   |   |-- helen
   |   |   |-- ibug
   |   |   |-- lfpw
   |   |   |-- Test
   |-- cofw
   |   |-- annotations
   |   |   |-- face_landmarks_cofw_test.csv
   |   |   |-- face_landmarks_cofw_train.csv
   |   |-- images
   |-- wflw
   |   |-- annotations
   |   |   |-- face_landmarks_wflw_test.csv
   |   |   |-- face_landmarks_wflw_test_blur.csv
   |   |   |-- face_landmarks_wflw_test_expression.csv
   |   |   |-- face_landmarks_wflw_test_illumination.csv
   |   |   |-- face_landmarks_wflw_test_largepose.csv
   |   |   |-- face_landmarks_wflw_test_makeup.csv
   |   |   |-- face_landmarks_wflw_test_occlusion.csv
   |   |   |-- face_landmarks_wflw_train.csv
   |   |-- images
   ...

You can directly download the native-dataset by baidu_OS key:3jnw. The repeated box csv files have more left,right,top,bottom, which is outputed by Retina-face network. You can choose use which one in experiments/*.json.

python tools/train.py --data_type=WFLW --dataset_exp=15 --model_type=PFLD_AVG --lr_reduce_patience=15 --model_dir=OCN --lr=0.01 --gpus=0 --batch_size=320 --loss_type=L1 --workers=8 --image_size=112

python tools/test_filelist.py --model_type=PFLD_AVG --resume_checkpoints={`PATH`} --data_type=WFLW --gpus=0

python tools/train_multi.py
--main_data=WFLW 
--aux_datas=300W 
--aux_ratios=3,1 
--ratios_decay=0.92 
--dataset_exp=15
--model_type=MULTI_FLAN2
--model_dir=ATF_W331  
--lr_reduce_patience=15
--lr=0.001
--gpus=0
--batch_size=300
--loss_type=L1
--workers=8
--auxdata_aug

python tools/test_filelist_MDT.py 
--model_type=MULTI_FLAN2
--resume_checkpoints=checkpoint/Mix/WFLW/MDT_WA71/29_checkpoint.pth
--main_data=WFLW
--gpus=0
--batch_size=2
--show_others
--aux_datas=AFLW

If you wanna to use the model to evaluate directly, you can download from baidu_os key: 7rqo.After that, you can run the above script with changed resume_checkpoints