Home

Awesome

ADNet: Leveraging Error-Bias Towards Normal Direction in Face Alignment

Paper link: ICCV 2021

Updated: 04/29/2022

Structure

FolderDescription
confConfigure files of different training tasks.
libCore implementation of all kinds of backbones, data loaders, losses and optimizers.
main.pyEntry script for training, testing or conversion.
trainer.pyEntry script for model training.
tester.pyEntry script for model testing.
evaluate.pyEvaluate script for model NME metrics calculation and visualization.
requirements.txtThe dependency list in python.

Dependencies

Local machine:

Resources

DatasetDatasetMetadata (train, test)Model (pytorch, onnx)
COFWofficialgoogle / baidugoogle / baidu
300Wofficialgoogle / baidugoogle / baidu
WFLWofficialgoogle / baidugoogle / baidu

Preparation

Training in local machine

python main.py --mode=train --config_name=alignment --device_ids=0,1,2,3

Testing in local machine

python main.py --mode=test --config_name=alignment --pretrained_weight=${model_path} --device_ids=0

Evaluation in local machine

python evaluate.py --mode=nme --config_name=alignment --model_path=${model_path} --metadata_path==${metadata_path} --image_dir=${image_dir} --device_ids=0

Framework

The framework of ADNet. ADNet

Performance

Table 1. Comparing with state-of-the-art methods on COFW by NME<sub>inter-pupils</sub>.

MethodNMEFR<sub>10%</sub>AUC<sub>10%</sub>
Human5.60--
RCPR8.5020.00-
TCDCN8.05--
DAC-CSR6.034.73-
Wu et al5.93--
Wing5.443.75-
DCFE5.277.290.3586
Awing4.940.990.6440
ADNet4.680.590.5317

Table 2. Comparing with state-of-the-art methods on 300W by NME<sub>inter-pupils</sub>.

MethodCommon SubsetChallenging SubsetFullset
PCD-CNN3.677.624.44
CPM+SBR3.287.584.10
SAN3.346.603.98
LAB2.985.193.49
DeCaFA2.935.263.39
DU-Net2.905.153.35
LUVLi2.765.163.23
AWing2.724.523.07
----------------------------------------------------
ADNet2.534.582.93

Table 3. Comparing with state-of-the-art methods on WFLW by NME<sub>inter-ocular</sub>.

MethodTestsetPose SubsetExpression SubsetIllumination SubsetMake-up SubesetOcclusion SubsetBlur Subset
ESR11.1325.8811.4710.4911.0513.7512.20
SDM10.2924.1011.459.329.3813.0311.28
CFSS9.0721.3610.098.308.7411.769.96
DVLN6.0811.546.785.735.987.336.88
LAB5.2710.245.515.235.156.796.12
Wing5.118.755.364.935.416.375.81
DeCaFA4.628.114.654.414.635.745.38
Awing4.367.384.584.324.275.194.96
LUVLi4.37------
ADNet4.146.964.384.094.055.064.79

Citation

@inproceedings{huang2021adnet,
  title={Adnet: Leveraging error-bias towards normal direction in face alignment},
  author={Huang, Yangyu and Yang, Hao and Li, Chong and Kim, Jongyoo and Wei, Fangyun},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={3080--3090},
  year={2021}
}

License

The project is released under the MIT License