liuwuliuyun/landmark

Facial Landmark Detection

Facial Landmark Detection

Network

Resnet-18

Env Install

Python 3.7.4 for cpu testing

Python 3.5.2 for gpu testing

# For cpu testing and evaluation
pip3 install ./requirements_cpu.txt
# For gpu training
pip3 install ./requirements_gpu.txt

Training and testing instructions

cd ROOT_DIR_OF_PRJ
# For training, model will save to ./weights
# All training arguments can be found at ./utils/args.py
python3 train.py --dataset_route path_to_dataset_root \
                 --dataset 'WFLW' \
                 --split 'train'
# For testing on cpu
python3 test_cpu.py --dataset_route path_to_dataset_root \
                    --dataset 'WFLW' \
                    --split 'test'
# For evaluation on cpu
python3 evaluation_cpu.py --dataset_route path_to_dataset_root \
                          --dataset 'WFLW' \ 
                          --split 'test'\ 
                          --eval_model_path = model_path \
                          --norm_way = 'face_size'

Training Dataset

WLFW Dateset

Architecture

1. Increase input size to 256*256
2. Remove avepooling layer
3. Change output size to 192 (x1, y1, ... , x98, y98)

Training Setup and testing results

Experiment 1

Training setup

Training epochs	Training batchsize	LR	Weight decay	Opt	Step Value
2000	4	2e-6	0	Adam	None

Testing result

Mean Error	Failure Rate	AUC	Normalized Way
13.80%	59.28%	0.1629	Face Size

Experiment 2

Training setup

Training epochs	Training batchsize	LR	Weight decay	Opt	Step Value
2700	4	1e-6	0	Adam	None

Testing result

Mean Error	Failure Rate	AUC	Normalized Way
14.04%	59.92%	0.1606	Face Size

Baseline (by Dongfeng Yu)

Training loss changed to log(1 + L2loss)

Training setup

Training epochs	Training batchsize	LR	Weight decay	Opt	Step Value
512	4	2e-6	0	Adam	None

Testing result

Mean Error	Failure Rate	AUC	Normalized Way
10.67%	38.56%	0.2735	Face Size

MobileNet-V2

Architecture

1. Increase input size to 256*256
2. Change output size to 192 (x1, y1, ... , x98, y98)

Training Setup and testing results

Experiment 1

Training loss changed to log(1 + L2loss)

Training setup

Training epochs	Training batchsize	LR	Weight decay	Opt	Step Value
992	4	2e-6	0	Adam	None

Testing result (mbv2_992.pth)

Mean Error	Failure Rate	AUC	Normalized Way
10.70%	44.48%	0.2299	Face Size

Testing result (mbv2_512.pth)

Mean Error	Failure Rate	AUC	Normalized Way
10.45%	42.00%	0.2427	Face Size

ShuffleNet-V2

Architecture

1. Increase input size to 256*256
2. Change output size to 192 (x1, y1, ... , x98, y98)

Training Setup and testing results

Experiment 1

Training loss changed to log(1 + L2loss)

Training setup

Training epochs	Training batchsize	LR	Weight decay	Opt	Step Value
700	4	2e-6	0	Adam	None

Testing result

Mean Error	Failure Rate	AUC	Normalized Way
10.83%	45.04%	0.2241	Face Size

TODO list

• Find out dataset defintion -- output sample images to 256*256 -- label landmarks and output heatmap (Finished 2019.12.26)
• Defind model and loss -- model backbone resnet-18 -- l2 loss (Finished 2019.12.30)
• Training and testing process -- with tensorboard (Finished 2020.01.02)
• Testing on 300w
• Quantilization model
• MNN Infer

References

1. Training and testing code are modified from Look_At_Boundary_PyTorch
2. Original paper and code 2018CVPR