Face-detection on IPC

[oberstet]

The face-detection demo is already packaged as a snap (#25).

The next step is do make it work it on the IPC327E we now have, publish a proper snap and document it (from a user point of view)

[oberstet]

actually, not sure about the status of https://github.com/crossbario/iotcookbook/tree/master/device/edge/face-detect @om26er ? I mean, could I connect a PSEye camera to a say intel NUC and get the face detect? IMO, would be also a very cool follow up demo in the style of https://www.katacoda.com/codelectron/scenarios/crossbar-demo ;) or some docker-compose/readme sugar to make it more approachable .. I could show/post it on mindsphere dev days already ..

[om26er]

I mean, could I connect a PSEye camera to a say intel NUC and get the face detect?

The current demo we have only processes images and not video streams. I will work with my PSEye camera tomorrow to get video facial detection on live videos.

From there we can extend the demo as we like.

[om26er]

@oberstet Well, I experimented with that just now.

Do we want to send the whole video stream from the front-end to the back-end, which will then detect faces on it and send it back to the front-end to draw ? Doing that there is going to be plenty of lag between the whats displayed on screen vs face rectangles.

[oberstet]

For the IPC setup, a setup more like this would be best:

• PS3 Eye == USB ==> IPC ==> Docker container A (with /dev/videoXX mounted)
• inside container A: OpenCV, Python cv2 to do two things: a) JPEG encode into (small batches of) video frames and WAMP-publish those to topic T1 and b) detect faces on raw image data and WAMP-publish detected rects/faces to topic T2
• inside container B: run a backend that listens for T2, and only in case a known face is recognized, it publishes to T3 (the backend actual does the face recognition on top of the face detection by matching against a database contained in the backend)
• Display <== HDMI == IPC <== Docker container/snap C: XBR dashboard, with access to video output (wayland), with a Tile1 that receives the raw video frames (jpeg) from topic T1, and a Tile2 that receives the rects/faces from topic T2, and a Tile3 that receives enriched detection events from T3

The parts in this setup are actually pretty complete to production setups I would get - at least that is the idea;) if you have more/other ideas, cool! lets discuss ..