Heatmap Regression via Randomized Rounding

---

[2020/09/02]: the paper is available on ArXiv.

[2020/10/28]: add new models for 106 facial landmarks.

[2021/09/21]: add new models for 68 facial landmarks (see models/README.md).

Introduction

This repo contains the facial landmark detection code for "Heatmap Regression via Randomized Rounding".

Demo

export PYTHONPATH=./:$PYTHONPATH
python examples/demo.py --image data/demo.jpg --model models/wflw/hrnet18_256x256_p2/

Test on WFLW (98 facial landmarks)

python examples/test_wflw.py --model models/wflw/hrnet18_256x256_p1/

Backbone	BBox	Resolution	#Params	FLOPs	NME (%)
HRNet-W18	P1	256x256	9.69M	4.84G	3.81
HRNet-W18	P2	256x256	9.69M	4.84G	3.95
MobileNetV2	P2	256x256	0.60M	0.51G	4.45
MobileNetV2	P2	160x160	0.60M	0.20G	4.58
MobileNetV2	P2	128x128	0.60M	0.13G	4.72

Test on LaPa (106 facial landmarks)

python examples/test_lapa.py --model models/lapa/hrnet18_256x256_p2/

Backbone	BBox	Resolution	#Params	FLOPs	NME (%)
HRNet-W18	P2	256x256	9.69M	4.86G	1.40
MobileNetV2	P2	256x256	0.60M	0.52G	1.69
MobileNetV2	P2	128x128	0.60M	0.13G	2.08

NOTE：All pretrained models can also be downloaded from google drive.

Citation

@article{yu2021heatmap,
  title={Heatmap Regression via Randomized Rounding},
  author={Yu, Baosheng and Tao, Dacheng},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2021}
}

Contact

Baosheng Yu, [email protected].

Acknowledgement

https://github.com/HRNet