Yolov2 tiny tensorflow for Movidius Neural Compute Stick (NCS)

This is a Tensorflow/NCSdk implementation of Yolov2. It was inpired by simo23 repo, see: https://github.com/simo23/tinyYOLOv2.

Credits to him for the functions to load the .weights file and functions to process the outcome of the Network.

I made this repo to make a tensorflow implementation of Yolov2 possible for the Movidius NCS, because until now I've only seen caffe implementations of Yolo for the NCS.

How to use it on the NCS

• clone the repo on a linux 16.04 system
• install ncsdk2
• run the following command in the terminal:

  python3 Script.py
  make compile
  python3 run.py -m [video/image] [-i PathToImage] [-v PathToVideo] [-v 0] #Last one is for using your webcam

• and done! The output of the network will look something like this:

How to use it normally / in debug mode

• clone the repo
• run the following script with python3:

  Script.py

• and done! The output should be the same as on the NCS

Note!

The processing power of the NCS is not as good as 4 titan X (pascal) GPU's, ofcourse. That is why you will get a processing time of ~6Hz / ~170ms, instead of 220Hz like advertised on the Yolov2 page. The biggest lost in preformance is made in conv7 and conv8:

If anybody can help me improve this lost, it would be much appreciated!! Any other sugestions and questions are welcome as well! You can contact me on this repo or at: verhaar.bastiaan@gmail.com