SPIGA: Shape Preserving Facial Landmarks with Graph Attention Networks.

This repository contains the models weights of SPIGA, a face alignment and headpose estimator that takes advantage of the complementary benefits from CNN and GNN architectures producing plausible face shapes in presence of strong appearance changes.

Setup

The repository is available on github

Results

WFLW Dataset

	NME_ioc	AUC_10	FR_10	NME_P90	NME_P95	NME_P99
full	4.060	60.558	2.080	6.766	8.199	13.071
pose	7.141	35.312	11.656	10.684	13.334	26.890
expression	4.457	57.968	2.229	7.023	8.148	22.388
illumination	4.004	61.311	1.576	6.528	7.919	11.090
makeup	3.809	62.237	1.456	6.320	8.289	11.564
occlusion	4.952	53.310	4.484	8.091	9.929	16.439
blur	4.650	55.310	2.199	7.311	8.693	14.421

MERLRAV Dataset

	NME_bbox	AUC_7	FR_7	NME_P90	NME_P95	NME_P99
full	1.509	78.474	0.052	2.163	2.468	3.456
frontal	1.616	76.964	0.091	2.246	2.572	3.621
half_profile	1.683	75.966	0.000	2.274	2.547	3.397
profile	1.191	82.990	0.000	1.735	2.042	2.878

300W Private Dataset

	NME_bbox	AUC_7	FR_7	NME_P90	NME_P95	NME_P99
full	2.031	71.011	0.167	2.788	3.078	3.838
indoor	2.035	70.959	0.333	2.726	3.007	3.712
outdoor	2.027	37.174	0.000	2.824	3.217	3.838

COFW68 Dataset

	NME_bbox	AUC_7	FR_7	NME_P90	NME_P95	NME_P99
full	2.517	64.050	0.000	3.439	4.066	5.558

300W Public Dataset

	NME_ioc	AUC_8	FR_8	NME_P90	NME_P95	NME_P99
full	2.994	62.726	0.726	4.667	5.436	7.320
common	2.587	44.201	0.000	3.710	4.083	5.215
challenge	4.662	42.449	3.704	6.626	7.390	10.095

BibTeX Citation

@inproceedings{Prados-Torreblanca_2022_BMVC,
  author    = {Andrés  Prados-Torreblanca and José M Buenaposada and Luis Baumela},
  title     = {Shape Preserving Facial Landmarks with Graph Attention Networks},
  booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
  publisher = {{BMVA} Press},
  year      = {2022},
  url       = {https://bmvc2022.mpi-inf.mpg.de/0155.pdf}
}