The diagram below shows the deployment of a vision pipeline / SAE instance.
A video stream is ingested through one or multiple video sources, which decode the stream and create
VideoFrame proto objects which are then serialized into base64 and pushed onto a Redis stream.
The object detector consumes these VideoFrame objects, detects objects on them, wraps the VideoFrame
into a DetectionOutput and pushes those onto another Redis stream. The object tracker consumes these
objects from Redis and tries to reassociate detected objects with previous observations, assigning an ID.
The result is pushed onto the Redis stream in the form of TrackingOutput objects.
The output is consumed by the database writer, which deserializes the messages and writes only the metadata
into a TimescaleDB (or a regular Postgres).
