hrabalm / video-search

video-search

Web application that scans video files and tags them using Deep Neural Networks and basic metadata extraction.

Video files can then be filtered by those tags.

Setup

Create a copy of docker-compose.local.example.yml called i.e. docker-compose.local.yml and modify volume mount binds suitably so that they map your multimedia folders into containers' /data/. Then run all the services by combining docker-compose.yml files with docker compose:

docker compose -f docker-compose.yml -f docker-compose.dev.yml -f docker-compose.local.yml build
docker compose -f docker-compose.yml -f docker-compose.dev.yml -f docker-compose.local.yml up

Be aware, that the images that are downloaded and built are quite large, around ~8GB and that building images can take quite long when run for the first time.

At the moment, docker files only setup development servers.

Technology stack

• React + MUI (frontend)
• Flask + Flask-RESTX (backend API)
• Dramatiq (task queue)
• MongoDB
• Tensorflow/Keras

Tagging

As far as semantic tags go, currently, video keyframes are passed through Keras EfficientNetV2-B0 and resulting classes with good enough confidence are considered as tags for the whole videos.

More interesting tagger is planned in the future.

Other than that, there is some support for simple metadata tags extracted from the video (e.g. codec, resolution, etc...)