SSH-TensorFlow

Introduction

This is an implementation of SSH: Single Stage Headless Face Detector reproduced using TensorFlow.

This code is modified from tf-faster-rcnn.

Prerequisites

1. You need a CUDA-compatible GPU to train the model.
2. You should first download WIDER Face for face detection.

Dependencies

• TensorFlow 1.4.1
• TF-Slim
• Python3.6
• Ubuntu 16.04
• Cuda 8.0

WIDER Face Eval result

SSH eval result without using an image pyramid between different backbones:

backbone	easy	medium	hard	training method
VGG16(origin paper)	0.919	0.907	0.814	--
VGG16	0.908	0.885	0.746	random training
VGG16	0.917	0.903	0.799	group training
ResNet50	0.902	0.880	0.689	random training
ResNet50	0.918	0.906	0.791	group training
MobileNetV1	0.897	0.874	0.720	random training
MobileNetV1	0.910	0.886	0.754	group training
MobileNetV2	0.865	0.820	0.576	random training

When the input image's shortest and longest side is limited as 400x800, the inference times are as follows:

backbone	inference time(GTX 1060)
VGG16	~56ms
ResNet50	~68ms
MobileNetV1	~23ms
MobileNetV2	~21ms

Note: About the difference between random training and group training, please see details in the script lib/model/train_val.py.

Result

demo_result(VGG16_SSH result):

Result on FDDB

VGG16-SSH:

Res50-SSH:

MobileNetV1-SSH:

Contents

1. Installation
2. Setup_data
3. Training
4. Evaluation
5. Demo
6. Models

Installation

• Clone the repository

git clone https://github.com/wanjinchang/SSH-TensorFlow.git

• Update your -arch in setup script to match your GPU

cd SSH-TensorFlow/lib
# Change the GPU architecture (-arch) if necessary
vim setup.py

GPU model	Architecture
TitanX (Maxwell/Pascal)	sm_52
GTX 960M	sm_50
GTX 1080 (Ti)	sm_61
Grid K520 (AWS g2.2xlarge)	sm_30
Tesla K80 (AWS p2.xlarge)	sm_37

Note: You are welcome to contribute the settings on your end if you have made the code work properly on other GPUs. Also even if you are only using CPU tensorflow, GPU based code (for NMS) will be used by default, so please set USE_GPU_NMS False to get the correct output.

• Build the Cython modules

make clean
make
cd ..

Setup_data

Generate your own annotation file from WIDER FACE dataset(eliminate the invalid data that x <=0 or y <=0 or w <=0 or h <= 0). One row for one image;
Row format: image_file_path box1 box2 ... boxN;
Box format: x_min y_min x_max y_max.
Here is an example:

0--Parade/0_Parade_marchingband_1_849 449.0 330.0 570.0 478.0
0--Parade/0_Parade_Parade_0_904 361.0 98.0 623.0 436.0
...

Or you can use my annotation files wider_face_train.txt and wider_face_val.txt under the folder data/ directly. And you should have a directory structure as follows:

data
   |--WIDER
         |--WIDER_train
             |--Annotations/
             |--images/ 
         |--WIDER_val
             |--Annotations/
             |--images/ 

Or you can follow the instructions of py-faster-rcnn here to setup WIDER Face datasets. The steps involve downloading data and optionally creating soft links in the data/ folder. If you find it is useful, the data/cache folder created on my side is also shared here.

Training

• Download pre-trained models and weights of backbones.The current code supports VGG16/ResNet_V1/MobileNet_Series models.

• Pre-trained models are provided by slim, you can get the pre-trained models from Google Driver or BaiduYun Driver. Uzip and place them in the folder data/imagenet_weights. For example, for VGG16 model, you also can set up like:

  mkdir -p data/imagenet_weights
  cd data/imagenet_weights
  wget -v http://download.tensorflow.org/models/vgg_16_2016_08_28.tar.gz
  tar -xzvf vgg_16_2016_08_28.tar.gz
  mv vgg_16.ckpt vgg16.ckpt
  cd ../..

  For ResNet101, you can set up like:

  mkdir -p data/imagenet_weights
  cd data/imagenet_weights
  wget -v http://download.tensorflow.org/models/resnet_v1_101_2016_08_28.tar.gz
  tar -xzvf resnet_v1_101_2016_08_28.tar.gz
  mv resnet_v1_101.ckpt res101.ckpt
  cd ../..

• Train

./experiments/scripts/train_ssh.sh [GPU_ID] [DATASET] [NET]
# GPU_ID is the GPU id you want to train on
# NET in {vgg16, res50, res101, res152, mobile, mobile_v2} is the backbone network arch to use
# DATASET {wider_face} is defined in train_ssh.sh
# Examples:
./experiments/scripts/train_ssh.sh 0 wider_face vgg16
./experiments/scripts/train_ssh.sh 1 wider_face res101

Note: Only support IMS_PER_BATCH=1 for training now, see details in the cfg files under the foder experiments/cfgs/.

By default, trained models are saved under:

output/[NET]/[DATASET]/default/

Test outputs are saved under:

output/[NET]/[DATASET]/default/[SNAPSHOT]/

Tensorboard information for train and validation is saved under:

tensorboard/[NET]/[DATASET]/default/
tensorboard/[NET]/[DATASET]/default_val/

Evaluation

Evaluation the trained models on wider face val dataset

./experiments/scripts/test_ssh.sh [GPU_ID] [DATASET] [NET]
# GPU_ID is the GPU you want to test on
# NET in {vgg16, res50, res101, res152} is the backbone network arch to use
# DATASET {wider_face} is defined in test_ssh.sh
# Examples: 
./experiments/scripts/test_ssh.sh 0 wider_face vgg16
./experiments/scripts/test_ssh.sh 0 wider_face mobile

Demo

• For ckpt demo

Download trained models from Models, then uzip to the folder output/, modify your path of trained model

# at repository root
GPU_ID=0
CUDA_VISIBLE_DEVICES=${GPU_ID} ./tools/demo.py.

or run python tools/demo.py directly.

• For frozen graph inference

Download the pb models(contained in Models) or frozen your model by yourself using script tools/convert_ckpt_to_pb.py, modify your path of trained model, then run python tools/inference.py.

Models

• vgg16_ssh(contains random_training and group training models) Google Driver, BaiduYun Driver
• res50_ssh(contains random_training and group training models) Google Driver, BaiduYun Driver
• mobile_ssh(contains random training and group training models/mobilenetv1_ssh) Google Driver, BaiduYun Driver
• mobilev2_ssh(contains random training/mobilenetv2_ssh) Google Driver, BaiduYun Driver

License

MIT LICENSE

TODOs

• Fix some bugs in training MobileNetV2_SSH
• Support multi-batch images training
• Multi-GPUs training
• Added facial landmarks localization, please see SSHA for details

References

1. SSH: Single Stage Headless Face Detector(https://arxiv.org/pdf/1708.03979.pdf). Mahyar Najibi, Pouya Samangouei, Rama Chellappa, Larry S. Davis.ICCV 2017.
2. tf-faster-rcnn
3. SSH(offical)