S3FD: Single Shot Scale-invariant Face Detector

Contents

1. Installation
2. Preparation
3. Train
4. Eval
5. Reference

Installation

1. Get the code. We will call the directory that you cloned Caffe into $CAFFE_ROOT

# 1. realize scale-compensation anchor matching strategy
# 2. realize random cropping square patches from original image 
git clone [email protected]:lippman1125/caffe_s3fd.git
cd caffe
git checkout ssd

Preparation

1. Download fully convolutional reduced (atrous) VGGNet.
   By default, we assume the model is stored in $CAFFE_ROOT/examples/s3fd/

2. Create the LMDB file.
```Shell
cd $CAFFE_ROOT
# Create the trainval.txt, test.txt, and test_name_size.txt in data/FACE/
./data/FACE/create_list.sh
# You can modify the parameters in create_data.sh if needed.
# It will create lmdb files for trainval and test with encoded original image:
#   - $HOME/data/faces_database/FACE/lmdb/FACE_trainval_lmdb
#   - $HOME/data/faces_database/FACE/lmdb/FACE_test_lmdb
# and make soft links at examples/VOC0712/
./data/FACE/create_data.sh

Train

1. Train your model .

./build/tools/caffe train --solver examples/s3fd/solver.prototxt  --gpu 1 --weights examples/s3fd/VGG_ILSVRC_16_layers_fc_reduced.caffemodel

Eval

1. ROC of FDDB compared with official, as follow:
   

2. ROC of FDDB compared with SSH/MTCNN, as follow:
   

3. examples
   
   

Reference

1. https://github.com/sfzhang15/SFD