FreeEnricher: Enriching Face Landmarks without Additional Cost

Paper link: AAAI 2023

Updated: 12/20/2023

Resources

Dataset	Dataset	Metadata (train, test)	Description
Enriched 300W	official	google / baidu	The metadata contains both train.tsv and test.tsv, where train.tsv is refined by FreeEnricher and test.tsv is manually labeled.
Enriched WFLW	official	google / baidu	The metadata only contains train.tsv, where train.tsv is refined by FreeEnricher.

Structure

Folder	Description
ADNet	The ADNet codebase of training and testing.
conf	The enriched version configure files.

Preparation

• Step1: Clone and Install ADNet.
• Step2: Replace the conf file of ADNet with enriched version.
• Step3: Download dataset and metadata of each dataset to data/alignment/${dataset} folder.
• Step4: Set the target dataset through configuring the ${data_definition} variable in conf/alignment.py script.
• Step5: Run the scripts in ADNet.

Framework

The framework of FreeEnricher.

Performance

Table 1. Performance of our method on enriched 300W.

Method	Network	NMEpoint	NMEedge
Baseline	ADNet + Line5	3.21	1.18
Ours	ADNet-FE5	3.06	0.98

Table 2. Comparing with state-of-the-art methods on original 300W and WFLW.

Method	300W	WFLW
LAB	3.49	5.27
HRNet	3.34	4.60
LUVLi	3.23	4.37
ADNet	2.93	4.14
ADNet-FE5	2.87	4.10

Citation

@inproceedings{huang2023freeenricher,
  title={FreeEnricher: enriching face landmarks without additional cost},
  author={Huang, Yangyu and Chen, Xi and Kim, Jongyoo and Yang, Hao and Li, Chong and Yang, Jiaolong and Chen, Dong},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={37},
  number={1},
  pages={962--970},
  year={2023}
}

License

The project is released under the MIT License