njcpe / facialrecognition Goto Github PK

The goal of this project is to detect Faces in live IP Camera footage. We use Tensorflow as our back end and build off of the SSD MobileNet Architecture. Versions using YOLOv3 and other Single Shot Detectors or Unified Detection Models are being explored for speed-accuracy trade-offs.

Video Exposition

We take in a live video stream from a network connected camera and perform image classification on the frames of this video. This video is then sent over LAN to an embedded device for display purposes (i.e. ARM Processor).

By taking this approach, we can allow users to explore image classification output on devices not powerful enough to perform these computations themselves. Additionally, any smartphone can act as the video source.

Built with

• TensorFlow
• OpenCV
• NVIDIA cuDNN
• COCO Dataset

• Make Install Guide for Linux (CPU Only)
• Make Install Guide for Linux (GPU)
• Reduce calls to frozen graph
• Rebuild to utilize Intel MKL-DNN, AVX2, JIT
• Extract frames from video at system edge
• Apply Transfer Learning and train using LFW Database

• Install Anaconda and add it to $PATH if you dont already have it.
• Install Tensorflow CPU and all dependencies:

  conda create --name tf python=3.5
  activate tf
  conda install scipy
  pip install tensorflow
  

• If everything is installed correctly then this test code

  >> import tensorflow as tf
  >> hello = tf.constant('Hello, TensorFlow!')
  >> sess = tf.Session()
  >> print(sess.run(hello))

  Should return

  Hello, TensorFlow!

Coming soon.

Setting Up TensorFlow GPU will take a long time, and there are no guarantees. The code will run much better on the GPU Version, however installation of TF-GPU on Windows is tedious at best and maddening at worst. Use this guide at your own risk.

• Ensure you have Microsoft C/C++ Build Tools 2015 installed. If you have Visual Studio 2017 installed, these tools are not installed by default. Make sure these are in your $PATH as well.
• Ensure that your NVIDIA Base Drivers are up to date. At time of writing, the current driver set is Release 390. Reboot after this step.
• Install the NVIDIA Cuda Toolkit. This must be version 9.0, version 9.1 is not yet supported. Add the bin folder and enclosing CUDA folder to $PATH.
• Install CuDNN version 7.0 and add that to $PATH.
• Reboot. If you got all of this installed correctly, the tedious part is over.
• Install Anaconda and add it to $PATH if you dont already have it.
• Install Tensorflow GPU and all dependencies:

  conda create --name tf-gpu python=3.5
  activate tf-gpu
  conda install scipy
  pip install tensorflow-gpu
  

• If everything is installed correctly then this test code

  >> import tensorflow as tf
  >> hello = tf.constant('Hello, TensorFlow!')
  >> sess = tf.Session()
  >> print(sess.run(hello))

  Should return

  Hello, TensorFlow!

• Go take a walk because you're done and you deserve it.

• Ensure you have the latest nvidia drivers, this can be determined using

nvidia-smi

• Install Tensorflow-GPU

pip install tensorflow-gpu

A full list of dependencies can be found in env.yml. Anaconda maintains that an equivalent enviroment to mine can be built using:

conda env create -f env.yml -n tf-gpu 

I have been unable to independenly verifiy this functionality. If anyone finds success with this method, let me know and I will update this to reflect any new information.

• If you get the following error:

   ModuleNotFoundError: No module named 'object_detection'

• Go to the repo parent directory (it should contain a folder 'object_detection') and execute the following commands:

   cd path/to/parent/dir/
   python setup.py setup
   python setup.py install

Inspired by TensorFlow Model Zoo

MIT License. Provided AS IS.

MIT © Noah Johnson