vre-video 0.2.7

Video Representations Extractor (VRE) Video reader

VRE Video

A python video reader that can read video frames using ffmpeg behind the scenes. Support for PIL/Numpy based directory of frames as well + in memory numpy frames for testing. Used in Video Representations Extractor (VRE) library to extract multiple modalities for multi-task learning (MTL) in machine learning from videos. However, the video reader container library can be used as-is and this repository is used to drive advancements/optimizations and supporting more video formats (thanks to ffmpeg). We have a dummy video player as well in the examples.

Install (pip):

pip install vre-video

Install (dev):

git clone https://gitlab.com/meehai/ffmpeg-video/
echo "$(pwd)/ffmpeg-video" >> ~/.bashrc
source ~/.bashrc
pip install -r ffmpeg-video/requirements.txt

Handle venv/conda/uv stuff on your own!

Usage:

from vre_video import VREVideo
video = VREVideo("video.mp4")
frame = video[ix] # returns a numpy array

Supports 3 backends for both reading and writing: numpy, Pillow and ffmpeg. It will auto-detect based on input: if a directory is provided it'll try to guess (png/jpg/npz/npy etc.) assuming it's a dir of frames (1.npz, ..., N.npz). If it's a path with suffix (i.e. .mp4, .mkv etc.) it will use the ffmpeg-based variant. Same for writing.

Support for youtube videos

Requires youtube-dl python package.

Usage:

from vre_video import VREVideo
video = VREVideo("https://www.youtube.com/...")
frame = video[ix] # returns a numpy array