Camera Motion Detection

This project uses optical flow to determine camera motion (pans and zooms).

NOTE: I'm not great at CV stuff so please excuse the probably not-ideal way of doing things here. And big thanks to Golan Levin and Alexander Porter for helping me figure it out...

Usage

detect.py will analyze a video frame by frame and record a csv file containing average camera motion angle, magnitude and "zoom factor." Zoomy-ness is determined by what percentage of pixels appear to be moving away from the center of the image.

To run:

python detect.py somevideo.mp4

If you have CUDA you can run the GPU optimized version with:

python detect.py somevideo.mp4 --gpu

You can also preview the analysis with:

python detect.py somevideo.mp4 --preview

Rendering out different camera motions

You can make use of the generated csv file in a few different ways. I've also included a render.py script that attempts to extract zooming and panning shots (your mileage may vary with it!).

To save out all the zoom shots:

python render.py somevideo.mp4 --zooms --output zooms.mp4

If you have mpv installed you can also preview this without rendering a new file:

python render.py somevideo.mp4 --zooms --preview

For panning shots, specify the desired panning angle like so:

python render.py somevideo.mp4 --pans --angle 180

Installation

CPU

Install the requirements with:

pip install -r requirements.txt

You're done!

GPU

If you've got an nvidia card you can use the GPU version which is significantly faster. You need opencv with cuda which you can either compile yourself or use a docker image.