This project is a pytorch implementation of a Single shot headless face detector. Original paper can be find here SSH face detector ICCV 2017 paper

official implementation is in Caffe SSH/mahyarnajibi

During my implementation of pytorch version, I referred the Faster-RCNN pytorch implementation jwyang/faster-rcnn.pytorch.

Compare to other faster-rcnn pytorch implementation, jwyang's implementation revised all layers from numpy to pytorch. I used anchor_target_layer , generate_anchor_box, propsal_layer and bbox layer for his implementation.

• Support python3
• Support pytorch-1.0
• Support distributed training (gradient average)
• Support multi-gpu(Multi-GPU can be done by using mulitiple distributed training on a single machine. The reason it is not support right now is because of the input size is not a fixed size.)
• Match mAP with original paper (current pretrain model have lower mAP (0.88)in wider face dataset that official implementation (0.907). Maybe more training is needed. Great thanks if anyone can help me find out what went wrong) i fixed a problem in anchor_target_layer. Now i can get( 0.905, 0.890, 0.809) vs SSH:(0.919,0.907,0.809)

1. Clone this repository
2. Create conda env by using pytorch.yml provide in this repository
3. Run make in the model directory:

cd model
make

For training on the WIDER dataset, you need to download the WIDER face training images and the face annotations from the dataset website. These files should be copied into data/datasets/wider/ (you can create symbolic links if you prefer to store the actual data somewhere else).

Before starting to train you should have a directory structure as follows:

data
  |--datasets
        |--wider
            |--WIDER_train/
            |--wider_face_split/
  |--imagenet_models
        |--VGG16.caffemodel

For training with the default parameters, you can call the train.py. As an example:

python3 train.py

For a list of all possible options run python3 train.py --help.

On each node, you can run train_dist.py For example

python3 train_dist.py --master_ip xxx.xxx.xxx.xxx --master_port xxxxx --rank 0 --world_size 2

--master_ip is your master node ip address and also --master_port. You need to make sure that each node would able to communicate with master node via TCP/IP. Rank 0 means master node, rank range from[0,world_size). World_size define total number of node you want to use.

You can treat a node with multi GPU as a multi node with single gpu. by doing this, we can utilize multi GPU on a single machine. For example

python train_dist.py --gpu='0' --rank=0 --world_size=2 --master_ip='127.0.0.1' --master_port=29500
python train_dist.py --gpu='1' --rank=1 --world_size=2 --master_ip='127.0.0.1' --master_port=29500

For test a image with trained model, you can call the test.py. As an example:

python3 test.py

For a list of all possible options run python3 train.py --help.

The evaluation on the WIDER dataset is based on the official WIDER evaluation tool which requires MATLAB. you need to download the validation images and the annotations (if not downloaded for training) from the WIDER dataset website. These files should be copied into the data/datasets/wider directory as follows:

data
  |--datasets
        |--wider
            |--WIDER_val/
            |--wider_face_split/

For evaluating with the default parameters, you can call the eval.py. As an example:

python3 eval.py

For a list of all possible options run python3 eval.py --help.

@misc{dechunwang_2018_1728203,
  author       = {dechunwang},
  title        = {dechunwang/SSH-pytorch: First release},
  month        = nov,
  year         = 2018,
  doi          = {10.5281/zenodo.1728202},
  url          = {https://doi.org/10.5281/zenodo.1728202}
}