fast-cv

This repository contains codes that show how video processing in Python can be significantly sped up as well as improved in quality

To run the tests on a video of your choice, run

./test.sh <video_name>

Files

• 1-traditional_cv.py - Performs video writing the normal way using OpenCV's VideoWriter class
• 2-multiproc_cv.py - Uses OpenCV's VideoWriter class but parallelizes the frames across available CPUs
• 3-multiproc_pipe_to_hdd_ffmpeg.py - Directly pipes the numpy arrays to an ffmpeg writer, allows video writing with modern codecs like HEVC and VP-9, etc without encoding, decoding from an intermediate codec and then re-encoding it
• 4-multiproc_pipe_to_stream_ffmpeg.py - Directly stores the parallelized stream in the memory avoiding storage write overhead at the cost of storing the part files in the RAM

Results

For a simple task like colour separation, we can see a speed up of 6.18x with 6 cores (on an i7-8700) and 2.35x (from 16 to 6 seconds on i5-4210U) with 2 cores

Presentation

The gitpitch presentation also in this repository only covers the methods in 2-multiproc_cv.py. The files indexed 3-... and 4-... combine piping of ffmpeg with the discussed multiprocessing methods