Heatmap In Heatmap

For models not using extra training data, Heatmap In Heatmap (HIH) has got:

• Rank 1 on WFLW Leaderboard
• Rank 1 on COFW Leaderboard
• Rank 3 on 300W Leaderboard.

Arxiv:HIH:Towards More Accurate Face Alignment via Heatmap in Heatmap

ICCVW:Revisting Quantization Error in Face Alignment

News

• [ April 18, 2022 ] We released HIH v2 in arxiv.

• [ April 17, 2022 ] Pretrained Model and evaluation code on WFLW dataset are released.

• [ March 22, 2022 ] HIH breaks the new records on WFLW and COFW.

• [ August 13, 2021 ] Accept by ICCV Workshop (Masked Face Recognition Challenge)

• [ April 03, 2021 ] We released HIH v1 in arxiv.

Introduction

This is the official code of HIH:Towards More Accurate Face Alignment via Heatmap in Heatmap. Compared with ICCVW version, we transform the subpixel regression problem into an interval classification problem and design a seamless loss to further improve performance. Moreover, we also adapt standard 4-stacked hourglass for experiments. We evaluate our methods on three datasets, COFW, WFLW and 300W.

Results

For inter-ocular NME, HIH reaches 4.08 on WFLW, 3.21 on COFW, 3.09 on 300W.

WFLW

COFW

300W

Installation

• Install Packages: pip install -r requirements.txt

• We have processed the dataset following PFLD practice, and you can download the training data and checkpoint directly at Baidu Drive (passwd:cjap) or Google Drive

• Unzip and move files into Best/WFLW and data/benchmark directory. Your folder structure should look like this

    ```
    HeatmapInHeatmap
    └───data
       │
       └───benchmark
       │   └───WFLW
       │   │
       │   └───COFW
       │   │     
       │   └───300W
    └───Best
       │
       └───WFLW
       │   └───WFLW.pth
       └───COFW
       │   └───COFW.pth
       └───300W
       │   └───300W.pth
    ```
  

Run Evaluation on WFLW

Evaluation cmd:

python tools/test_all.py --config_file experiments/Data_WFLW/HIHC_64x8_hg_l2.py --resume_checkpoint Best/WFLW/WFLW.pth

Future Plans

• Release evaluation code and pretrained model on WFLW dataset.

• Release training code on WFLW dataset.

• Release pretrained model and code on 300W and COFW dataset.

• Release facial landmark detection API

Citations

If you find this work or code is helpful in your research, please cite the following papers.

@inproceedings{DBLP:conf/iccvw/LanHC21,
  author    = {Xing Lan and
               Qinghao Hu and
               Jian Cheng},
  title     = {Revisting Quantization Error in Face Alignment},
  booktitle = {{IEEE/CVF} International Conference on Computer Vision Workshops,
               {ICCVW} 2021, Montreal, BC, Canada, October 11-17, 2021},
  pages     = {1521--1530},
  publisher = {{IEEE}},
  year      = {2021},
  url       = {https://doi.org/10.1109/ICCVW54120.2021.00177},
  doi       = {10.1109/ICCVW54120.2021.00177},
  timestamp = {Wed, 06 Apr 2022 11:41:39 +0200},
  biburl    = {https://dblp.org/rec/conf/iccvw/LanHC21.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

@article{DBLP:journals/corr/abs-2104-03100,
  author    = {Xing Lan and
               Qinghao Hu and
               Jian Cheng},
  title     = {{HIH:} Towards More Accurate Face Alignment via Heatmap in Heatmap},
  journal   = {CoRR},
  volume    = {abs/2104.03100},
  year      = {2021},
  url       = {https://arxiv.org/abs/2104.03100},
  eprinttype = {arXiv},
  eprint    = {2104.03100},
  timestamp = {Wed, 06 Apr 2022 11:41:43 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2104-03100.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

Acknowledgments

This repository borrows or partially modifies hourglass model and data processing code from Hourglass and HRNet

License

This repository is released under the Apache 2.0 license.