Resolution

[highaltidude]

So if models are in a 300 x 300px images, why does having a higher resolution offer no performance increase? Let's say I have a 4k camera, there should be a huge canvas for analytics to find objects. If Frigate reduces the area to 300x300 doesn't that make the detect range of Frigate extremely limited? Love the project. Thanks!

[blakeblackshear]

A higher resolution model would in theory offer better performance, but likely wont fit on the Coral or be performant enough for realtime analysis. These aren't frigate-specific models. They are provided by Google, and they are 300x300 or 320x320.

Let's say I have a 4k camera, there should be a huge canvas for analytics to find objects

This is why frigate doesn't pass the entire frame to object detection. It looks for motion and then creates a bounding box around a subset of the frame and sends that for object detection. That bounding box area is always resized to the model size (320x320). When running detection on a 4k stream, the bounding box for motion when a person is near will likely be much larger than 320x320. Resize operations are always a balance of speed and data loss, and frigate prioritizes speed, so the more you shrink the image, the worse the image quality. Of course if you are using a 4k stream to detect very small, far away objects, this won't be a problem.

[poweredgenl]

so when i understand you correctly - het goal is to balance between res and distance of the person on screen... eg if people where to be further away you need the more pixels to get to a suitable 300x300 model? so if ppl would be further away on the camera ... a higher resolution is more beneficial?

[blakeblackshear]

Yes

[poweredgenl]

ok clear - then im getting the hang of it - in my case with wide view security cams - a higher reso would be better - eg limit fps to not overstretch the RPI... and then in MQTT as i use frigate with compreface as well increase the snapshot it will sent to MQTT ? would this be also your approach?

[blakeblackshear]

In general, I don't consider MQTT to be a great way to get an image for further processing. You can't reliability connect the snapshot to an event id. I would use the events mqtt topic and fetch images from the http api as needed for further processing.

[haggercody]

i got a coral so i could test out using my main stream for detect because our camera is pretty wide angle and resolution really drops when you get just a bit away from our house so we've never been able to motion detect there with blue iris without getting a ton of noise. Frigate is doing an awesome job on this. The only downside i've noticed is that it appears there is an ffmpeg process that is running a bit more cpu than before--my guess is the detect stream. seems to hover about 20% on one cpu.