ORFormer (WACV 2025 Oral)

This is the offcial repository of WACV 2025 Oral paper ORFormer: Occlusion-Robust Transformer for Accurate Facial Landmark Detection.

Project Website

arxiv Paper

WACV Paper

Architecture

Quantitative Results

Training Commands

All experiments can be run on a NVIDIA GTX1080Ti(11Gb RAM).

Quantized Heatmap Generator

#WFLW
CUDA_VISIBLE_DEVICES=x python train_heatmap_generator.py --dataset WFLW --name [run_name] --resultDir [result_dir] --lr 0.0007 --batch_size 128 --alpha 100
#300W
CUDA_VISIBLE_DEVICES=x python train_heatmap_generator.py --dataset 300W --name [run_name] --resultDir [result_dir] --lr 0.0008 --batch_size 128 --alpha 1000
#COFW
CUDA_VISIBLE_DEVICES=x python train_heatmap_generator.py --dataset COFW --name [run_name] --resultDir [result_dir] --lr 0.0007 --batch_size 128 --alpha 100

ORFormer

#WFLW
CUDA_VISIBLE_DEVICES=x python train_ORFormer.py --dataset WFLW --name [run_name] --resultDir [result_dir] --lr 0.0001 --batch_size 64 --alpha 50 --vit ORFormer
#300W
CUDA_VISIBLE_DEVICES=x python train_ORFormer.py --dataset 300W --name [run_name] --resultDir [result_dir] --lr 0.0001 --batch_size 64 --alpha 100 --vit ORFormer
#COFW
CUDA_VISIBLE_DEVICES=x python train_ORFormer.py --dataset COFW --name [run_name] --resultDir [result_dir] --lr 0.0001 --batch_size 64 --alpha 50 --vit ORFormer

Integration with FLD Methods

#WFLW
CUDA_VISIBLE_DEVICES=x python train_HGNet_with_ORFormer.py --dataset WFLW --name [run_name] --resultDir [result_dir] --lr 0.001 --batch_size 16 --alpha 0.05 --heatmap ORFormer
#300W
CUDA_VISIBLE_DEVICES=x python train_HGNet_with_ORFormer.py --dataset 300W --name [run_name] --resultDir [result_dir] --lr 0.001 --batch_size 16 --alpha 0.05 --heatmap ORFormer
#COFW
CUDA_VISIBLE_DEVICES=x python train_HGNet_with_ORFormer.py --dataset COFW --name [run_name] --resultDir [result_dir] --lr 0.001 --batch_size 16 --alpha 0.05 --heatmap ORFormer

Performance

	WFLW	300W	COFW	Training	Inference
Quantized Heatmap Generator	L2 Loss: 26.72	L2 Loss: 14.12	L2 Loss: 30.32	train_heatmap_generator.py	test_heatmap_generator.py
ORFormer	L2 Loss: 20.22	L2 Loss: 10.97	L2 Loss: 23.06	train_ORFormer.py	test_ORFormer.py
Integration	NME Loss: 3.86	NME Loss: 2.90	NME Loss: 4.46	train_HGNet_with_ORFormer.py	test_HGNet_with_ORFormer.py

Links

Datasets

weights

Citation

@InProceedings{Chiang_2025_WACV,
    author    = {Chiang, Jui-Che and Hu, Hou-Ning and Hou, Bo-Syuan and Tseng, Chia-Yu and Liu, Yu-Lun and Chen, Min-Hung and Lin, Yen-Yu},
    title     = {ORFormer: Occlusion-Robust Transformer for Accurate Facial Landmark Detection},
    booktitle = {Proceedings of the Winter Conference on Applications of Computer Vision (WACV)},
    month     = {February},
    year      = {2025},
    pages     = {784-793}
}

Acknowledgement

This repository is build on top of SLPT, ADNet, STAR and vit_pytorch. Huge appreciation for the excellent codebases provided by these projects!