Skip to content

README.md

Latest commit

 

History

History
39 lines (31 loc) · 2.79 KB

README.md

File metadata and controls

39 lines (31 loc) · 2.79 KB

Prerequisites

  • Python 3.10
  • Make sure that pip can access the correct github repositories (visionapi and visionlib) by running source set_local_git_creds.sh and supplying your username and an access token, which has read access to vision-api and vision-lib. On some systems Git seems to be able to pick up ssh keys automatically...
  • Create and activate python virtualenv 
    python3 -m venv .venv
source .venv/bin/activate
  • Install dependencies: pip install -r requirements.txt
  • Install libturbojpeg on your OS (e.g. apt install libturbojpeg)

Tools

Visual Introspection (watch.py)

The watch.py script can be used to visually look into the data flows within the pipeline. On a technical level, it attaches to a Redis stream of your choice and then tries to guess from the name prefix which stage output it has to decode. It will then render (and annotate, if possible) every output object / proto it receives from Redis. You can exit the program by pressing q.

Examples

  • python watch.py displays a menu with all available streams for ease of use (and after selection renders content of that stream)
  • python watch.py -h shows all available options
  • python watch.py -s objectdetector:video1 renders frames with detected objects (assuming that objectdetector:* contains outputs of the objectdetector stage, which is default)

Caveats

  • The annotations are not exactly pretty anymore, because I wanted to remove the dependency on the ultralytics library (which has great plotting features) -> you are more than welcome to improve that :)
  • Data transfer from Redis and rendering will increase your system load by another few percent

Pipeline Recording (record.py)

The record.py script provides a simple way to record messages from some or all Redis streams into a file, i.e. create a log of all pipeline activities / state. See python record.py -h for how to use it.

Pipeline Playback (play.py)

The play.py script plays back a pipeline log into a running pipeline (i.e. at least a running Redis instance). It'll read the log file it is given and play back all messages into the corresponding streams they were recorded from. The messages will be spaced exactly as they were recorded (i.e. a 5fps recording will be played back at the same speed). See python play.py -h for how to use it.

JSON Output (echo.py)

The echo.py script echoes all SAE messages it receives into stdout as a JSON string (output of protobufs MessageToJSON()). Frame data is removed by default as to not clutter the output. echo.py can be very useful when combined with other tools like jq. For example, to print the source id, frame timestamp and number of detections for each received message: python echo.py | jq -r '[.frame.sourceId, .frame.timestampUtcMs, (.detections | length)] | @tsv'