LPR does not seem to improve/adapt detection based on the entire length of an event clip

[tetele]

When my car is in it's parking spot, the height of the security camera makes reading its license plate difficult, as the top of the plate is obstructed by the rear bumper. Additionally, due to the angle at which I park my car, the first frames in which the license plate is detected are also useless for actually recognizing the license number.

However, during the parking maneuver, there are frames where the license plate is very clearly detectable and readable. Despite this, Frigate doesn't seem to care about getting the best possible frame for reading the license plate for my (persistent object) car. If I haven't properly explained, I can provide a sample video, but I'd rather do it privately.

Is the behavior I'm describing intended? Am I doing something wrong?

[hawkeye217]

There are a number of factors involved with your situation, and it's hard to make a specific recommendation without seeing your config and the actual scene.

I would begin by enabling the LPR debug logs so you can see exactly what Frigate is doing.

Follow the debugging steps in the beta documentation: https://deploy-preview-16390--frigate-docs.netlify.app/configuration/license_plate_recognition#how-do-i-debug-lpr-issues

[tetele]

Thanks! I'll give that a go and report back.

[tetele]

The large model doesn't seem to change too much. I'll run some more tests tomorrow.

However, it's becoming apparent that I need a bigger res stream to analyze, which kinda conflicts with the instruction to use lower res streams for object detection. Any advice on that?

[hawkeye217]

Generally lower detect resolutions are recommended for object detection, but since every scene is different, you should just select a suitable resolution for the stream in your camera's firmware that fits your specific scene and requirements. Slowly increase it until you find a value that works best for recognition and still maintains a reasonable CPU usage.

[tetele]

I've quadrulped the detect resolution (from 1280x720 to 2560x1440) and the result is dramatic in terms of recognition quality. I see some CPU impact, but I can afford the extra usage. As a desirable side effect, face recognition gets much better training data.

Thanks again for the tips!

[hawkeye217]

Great, glad it's working for you now. Happy to help.