This repository includes the training and evaluating codes for SFA face detector, implemented in pycaffe. The SFA codes is an extension of the SSH repository. Our method is trained on the training set of the WIDER FACE dataset. If it is useful, please light up a star for this repository.
1.1. Clone the repository:
git clone --recursive https://github.com/shiluo1990/SFA.git
1.2. Install cuDNN and NCCL (used for multi-GPU training).
1.3. Caffe and pycaffe: You need to compile the caffe-sfa repository which is a Caffe fork compatible with SFA. Caffe should be built with cuDNN, NCCL, and python layer support (set by default in Makefile.config.example). You also need to make pycaffe.
1.4. Install python requirements:
pip install -r requirements.txt
1.5. Run make in the lib directory:
cd lib
make
All SFA default settings and configurations are saved in SFA/configs/default_config.yml. More details can be seen in SFA/configs/default_config.yml as an example config files.
We provide 5 configuration files in SFA/configs/ to perform MS-Training and MS-Testing in single_scale, 4_scale, and wide_scale.
single_scale 4_scale wide_scale
training wider_single_scale_training.yml wider_4_scale_training.yml
testing wider_single_scale_testing.yml wider_4_scale_testing.yml wider_wide_scale_testing.yml
They can be overwritten by passing an external configuration file to the module --cfg [path-to-config-file].
Note that you can adjust testing scales to flexibly trade off detection accuracy and efficiency for real applications. To avoid out of GPU memory, you can properly decrease the max scale in MS-Testing.
Solver parameters are saved in SFA/models/solver_sfa.prototxt (e.g. learning rate, gamma, stepsize, momentum, etc.).
SFA training and testing network are described in
SFA/models/train_sfa.prototxt
SFA/models/test_sfa.prototxt
To run the demo, you need to download the provided pre-trained SFA model from Baidu Yun via the following link and extraction code :
link: https://pan.baidu.com/s/1LAWhtiATbPHKweMeY8hVaQ
extraction code: ica9
or Google Drive
link: https://drive.google.com/drive/folders/1W9U1gYcRDMpIEH5DG8JloqZNOU89c5GR?usp=sharing
By default, the model is saved into a folder named data/SFA_models.
Before starting to run demo, you should have a directory structure as follows:
data
|--demo
|--demo.jpg
|--SFA_models
|--SFA.caffemodel
To train your own SFA model, you also need to download the VGG-16 ImageNet model as pre-trained model. The following script downloads the model into the default directory:
bash scripts/download_imgnet_model.sh
If you can't download the pre-trained VGG-16 ImageNet model via the above script, please try to download from Baidu Yun via the following link and extraction code :
link: https://pan.baidu.com/s/1VqiwWHiFPnDefMymgdgbdA
extraction code: 7790
By default, the model is saved into a folder named data/imagenet_models.
For training and evaluation on the WIDER dataset, you need to download the WIDER FACE dataset from the WIDER FACE dataset website. These files should be copied into data/datasets/wider/.
Before starting to train and evaluate, you should have a directory structure as follows:
data
|--datasets
|--wider
|--WIDER_train/
|--WIDER_val/
|--WIDER_test/
|--wider_face_split/
|--imagenet_models
|--VGG16.caffemodel
After downloading the SFA model, you can run the demo with the default configuration as follows:
python demo.py
If everything goes well, the following detections should be saved as data/demo/demo_detections_SFA.png.
For a list of possible options run: python demo.py --help
More detection results in our paper are offered in paper_result/detection_result.
For training with the default parameters, you can call the main_train module with a list of GPU ids. As an example:
python main_train.py --gpus 0,1
For a list of all possible options run python main_train.py --help.
By default, the models are saved into the output/[EXP_DIR]/[db_name]/ folder (EXP_DIR is set to sfa by default and can be changed through the configuration files,
and db_name would be wider_train in this case).
The evaluation on the WIDER dataset is based on the official WIDER evaluation tool which requires MATLAB.
We performed evaluation with the default configuration by calling the main_test module:
python main_test.py --model [path-to-the-trained-model]
For a list of possible options run python main_test.py --help.
The evaluation outputs are saved into output/[EXP_DIR]/[db_name]/[net_name] (EXP_DIR is set to sfa by default and can be changed by passing a config file, net_name can be directly passed to the module and is set to SFA by default, and db_name would be wider_val in this case). This includes the detections saved as text files in a folder named detections,detections saved as a pickle file, and the WIDER evaluation plots saved in a folder named wider_plots.
Please note that the detections will be cached by default and will not be re-computed again (the caching can be disabled by passing the --no_cache argument.)
The evaluation results in our paper are offered in paper_result/evaluation.
If this project helps your research, please consider citing the following papers:
@article{luo2019sfa,
title={SFA: Small faces attention face detector},
author={Luo, Shi and Li, Xiongfei and Zhu, Rui and Zhang, Xiaoli},
journal={IEEE Access},
volume={7},
pages={171609--171620},
year={2019},
publisher={IEEE}
}
@inproceedings{najibi2017ssh,
title={{SSH}: Single Stage Headless Face Detector},
author={Najibi, Mahyar and Samangouei, Pouya and Chellappa, Rama and Davis, Larry},
booktitle={The IEEE International Conference on Computer Vision (ICCV)},
year={2017}
}
To ask questions or report issues, please open an issue on the issues tracker.
Email: 1093945458@qq.com or shiluo1990@hotmail.com
Sree Charan：Upload SFA.caffemodel on Google Drive.