Generating the middle frames of a video with TensorFlow Hub

Another interesting application of deep learning to videos involves frame generation. A fun and practical example of this technique is slow motion, where a network decides, based on the context, how to create intervening frames, thus expanding the length of a video and creating the illusion it was recorded with a high-speed camera (if you want to read more about it, refer to the See also… section).

In this recipe, we'll use a 3D convolutional network to produce the middle frames of a video, given only its first and last frames.

For this purpose, we'll rely on TFHub.

Let's start this recipe.

Getting ready

We must install TFHub and TensorFlow Datasets:

$> pip install tensorflow-hub tensorflow-datasets

The model we'll use was trained on the BAIR Robot Pushing Videos dataset, which is available in TensorFlow Datasets. However, if we access it through the library, we'll...