[Support]: Compare detection param on the same go2rtc stream

[xavan83]

Describe the problem you are having

I setup two camera config against the same go2rtc stream (mainly cars crossing in front of the camera).
I want to test the parameter that I can change through mqtt (threshold,contour,contrast).
I made some tests and using the same value for these 3params and 60% of the time events are the same for both config.
40% of the time events are only detected on one of the config.
If I take the clip of a detected event to test against the undetected config, this time the event is detected.
The only thing that happen when I test against a mp4 file is that the accuracy can vary from 75% to 90% for the same clip running in loop.
Can the hardware be to limited for 1go2rtc stream and 2cameras config? Celeron J4105 and 4Gb ram, dual coral pci express ssd 1Gb (Frigate+ model)

Version

13.2

Frigate config file

sonnettetest:
    ffmpeg:
      inputs:
        - path: /testcamera/pastrouvesursensible1.mp4
          input_args: -re -stream_loop -1 -fflags +genpts
          roles:
            - detect
    detect:
      fps: 5
      enabled: True
      width: 1280
      height: 720
    motion:
      mask:
        - 638,463,1280,422,1280,0,0,0,0,458
      contour_area: 10 
      threshold: 30 
      improve_contrast: off
sonnette:
    ffmpeg:
      inputs:
        - path: /testcamera/pastrouvesursensible1.mp4
          input_args: -re -stream_loop -1 -fflags +genpts
          roles:
            - detect
    detect:
      fps: 5
      enabled: True
      width: 1280
      height: 720
    motion:
      mask:
        - 638,463,1280,422,1280,0,0,0,0,458
      contour_area: 10 
      threshold: 30 
      improve_contrast: off

Relevant log output

2024-03-21 12:37:46.880394255  [INFO] Preparing Frigate...
2024-03-21 12:37:46.905590599  [INFO] Starting Frigate...
2024-03-21 12:37:49.126303181  [2024-03-21 12:37:49] frigate.app                    INFO    : Starting Frigate (0.13.2-6476f8a)
2024-03-21 12:37:49.446681714  [2024-03-21 12:37:49] peewee_migrate.logs            INFO    : Starting migrations
2024-03-21 12:37:49.455755658  [2024-03-21 12:37:49] peewee_migrate.logs            INFO    : There is nothing to migrate
2024-03-21 12:37:49.466462804  [2024-03-21 12:37:49] frigate.app                    INFO    : Recording process started: 298
2024-03-21 12:37:49.471639114  [2024-03-21 12:37:49] frigate.app                    INFO    : go2rtc process pid: 89
2024-03-21 12:37:49.534736911  [2024-03-21 12:37:49] detector.coral1                INFO    : Starting detection process: 308
2024-03-21 12:37:49.560124663  [2024-03-21 12:37:49] frigate.detectors.plugins.edgetpu_tfl INFO    : Attempting to load TPU as pci:1
2024-03-21 12:37:49.561482566  [2024-03-21 12:37:49] frigate.app                    INFO    : Output process started: 317
2024-03-21 12:37:49.580288626  [2024-03-21 12:37:49] detector.coral2                INFO    : Starting detection process: 310
2024-03-21 12:37:49.580296594  [2024-03-21 12:37:49] frigate.detectors.plugins.edgetpu_tfl INFO    : Attempting to load TPU as pci:0
2024-03-21 12:37:49.581297907  [2024-03-21 12:37:49] frigate.detectors.plugins.edgetpu_tfl INFO    : TPU found
2024-03-21 12:37:49.592017507  [2024-03-21 12:37:49] frigate.detectors.plugins.edgetpu_tfl INFO    : TPU found
2024-03-21 12:37:49.661435256  [2024-03-21 12:37:49] frigate.app                    INFO    : Camera processor started for sonnettenight: 351
2024-03-21 12:37:49.674187819  [2024-03-21 12:37:49] frigate.app                    INFO    : Camera processor started for sonnette: 352

FFprobe output from your camera

-

Frigate stats

No response

Operating system

Other Linux

Install method

Docker Compose

Coral version

M.2

Network connection

Wired

Camera make and model

Hikvision

Any other information that may be helpful

No response

[NickM-27]

There are a few things to clear up:

1. detection results are not deterministic. This means that it is not expected that rerunning on a clip will have the same detection results apply every time.
2. The main goal of motion detection should be watching motion boxes and only having visible motion detected. Using events will likely lead to large swinging changes that are more difficult to fine tune.

[xavan83]

I don't understand your second point.
Event is the final result i.e.: A car is match with 97% by the first config (true positive). The second with same param does not match it.
I can see that both config detect the motion and process detection.
I just want to know if this is not the result of a hardware overload that cause the second detection process to stop the execution (however the skipped frame value is 0).

[NickM-27]

My point is that the goal of adjusting motion parameters should not be to affect what events get created. It should be to have motion boxes detected under expected circumstances only.

If skipped fps is 0 then the detector was not saturated.

[xavan83]

Does the motion configuration has an impact on the initial box send to the detector? May be if the box is far away for the target object the detection process is stopped before reaching the minimum score?
I will try to lower the min_score to see if this changes something.
Can I add some more logs to debug this behavior?

[NickM-27]

yes, the areas of motion detection will affect the region sent to the detector