Repository files navigation

Dockerface is a deep learning face detector. It deploys a trained Faster R-CNN network on Caffe through an easy to use docker image. Bring your videos and images, run dockerface and obtain videos and images with bounding boxes of face detections and an easy to use face detection annotation text file.

The docker image is large for now because OpenCV has to be compiled and stored in the image to be able to use video and it takes up a lot of space.

Technical details and some experiments are described in theArxiv Tech Report.

If you find Dockerface useful in your research please consider citing:

@ARTICLE{2017arXiv170804370R,   author = {{Ruiz}, N. and {Rehg}, J.~M.},    title = "{Dockerface: an easy to install and use Faster R-CNN face detector in a Docker container}",  journal = {ArXiv e-prints},archivePrefix = "arXiv",   eprint = {1708.04370}, primaryClass = "cs.CV", keywords = {Computer Science - Computer Vision and Pattern Recognition},     year = 2017,    month = aug,   adsurl = {http://adsabs.harvard.edu/abs/2017arXiv170804370R},  adsnote = {Provided by the SAO/NASA Astrophysics Data System}}

Install NVIDIA CUDA (8 - preferably) and cuDNN (v5 - preferably)

https://developer.nvidia.com/cuda-downloadshttps://developer.nvidia.com/cudnn

Install docker

https://docs.docker.com/engine/installation/

Install nvidia-docker

wget -P /tmp https://github.com/NVIDIA/nvidia-docker/releases/download/v1.0.1/nvidia-docker_1.0.1-1_amd64.debsudo dpkg -i /tmp/nvidia-docker*.deb && rm /tmp/nvidia-docker*.deb

Go to your working folder and create a directory called data,your videos and images should go here. Also create a folder called output.

cd $WORKING_DIRmkdir datamkdir output

Run the docker container

sudo nvidia-docker run -it -v $PWD/data:/opt/py-faster-rcnn/edata -v $PWD/output/video:/opt/py-faster-rcnn/output/video -v $PWD/output/images:/opt/py-faster-rcnn/output/images natanielruiz/dockerface:latest

Now we have to recompile Caffe for it to work on your own machine.

cd caffe-fast-rcnnrm -rf buildmkdir buildcd buildcmake -DUSE_CUDNN=1 ..make -j20 && make pycaffecd ../..

Finally use this command toprocess a video

python tools/run_face_detection_on_video.py --gpu 0 --video edata/YOUR_VIDEO_FILENAME --output_string STRING_TO_BE_APPENDED_TO_OUTPUTFILE_NAME --conf_thresh CONFIDENCE_THRESHOLD_FOR_DETECTIONS

Use this command toprocess an image

python tools/run_face_detection_on_image.py --gpu 0 --image edata/YOUR_IMAGE_FILENAME --output_string STRING_TO_BE_APPENDED_TO_OUTPUTFILE_NAME --conf_thresh CONFIDENCE_THRESHOLD_FOR_DETECTIONS

Also if you are looking to convenientlyprocess all images in one folder use this command

python tools/facedetection_images.py --gpu 0 --image_folder edata/IMAGE_FOLDER_NAME --output_folder OUTPUT_FOLDER_PATH --conf_thresh CONFIDENCE_THRESHOLD_FOR_DETECTIONS

The default confidence threshold is 0.85 which works for high quality videos or images where the faces are clearly visible. You can play around with this value.

The columns contained in the output text files are:

Forvideos:

frame_number x_min y_min x_max y_max confidence_score

Forimages:

image_path x_min y_min x_max y_max confidence_score

Where (x_min,y_min) denote the coordinates of the upper-left corner of the bounding box in image intrinsic coordinates and (x_max, y_max) denote the coordinates of the lower-right corner of the bounding box in image intrinsic coordinates. (ref.https://www.mathworks.com/help/images/image-coordinate-systems.html)confidence_score denotes the probability output of the model that the detection is correct (it is a number included in [0,1])

Voila, that easy!

After you're done with the docker container you can exit.

exit

You want to restart and re-attach to this same docker container so as to avoid compiling Caffe again. To do this first get the id for that container.

sudo docker ps -a

It should be the last one that was launched. Take note of CONTAINER ID. Then start and attach to that container.

sudo docker start CONTAINER_IDsudo docker attach CONTAINER_ID

You can now continue processing videos.

Nataniel Ruiz and James M. Rehg
Georgia Institute of Technology

Credits:Originaldockerface logo made byFreepik fromFlaticon is licensed byCreative Commons BY 3.0, modified by Nataniel Ruiz.