TensorFlow Tutorial - Video Data
Notes by Magnus Erik Hvass Pedersen: https://github.com/Hvass-Labs/TensorFlow-Tutorials/blob/master/09_Video_...
The previous Tutorial #08 showed how to use the pre-trained Inception model on the CIFAR-10 data-set in so-called Transfer Learning. This tutorial shows how to use your own images.
For demonstration, we use a new data-set called Knifey-Spoony that contains thousands of images of cutlery knives, spoons and forks on a few different backgrounds. The training-set contains 4170 images and the test-set contains 530 images. The classes are named knifey, spoony and forky as a reference to The Simpsons.
The images in the knifey-spoony data-set were created from video-files using a small Python script that is run on Linux (it requires the avconv-program for conversion from videos to images). This allows you to easily create very large data-sets with thousands of images from just a few minutes of video recordings.
This tutorial builds on the previous tutorials so you should be familiar with Tutorial #08 on Transfer Learning, as well as earlier tutorials on how to build and train Neural Networks in TensorFlow.