Installing Coral hasn't really improved Frigate event captures

With my old setup using just the NUC CPU I would get false positives when, at night, there was fog or during the day with cloud reflections in the car glass. I’m still getting false positives after installing the Coral USB Accelerator. I’m wondering if my Frigate config needs to be modified. I have HA setup in a Proxmox VM.

Here’s my Frigate config. Do I have something missing in this? Any hints really appreciated.

mqtt:
  host: xxx.x.x.xxx
  user: Tom
  password: xxxxxxx

detectors:
  coral:
    type: edgetpu
    device: usb

logger:
  default: info
  logs:
    frigate.record: debug
    
cameras:
  driveway_cam:
    ffmpeg:
      inputs:
        - path: rtsp://admin:xxxxxxxxxxxxxxxxxx/cam/realmonitor?channel=1&subtype=0&authbasic=64
          roles:
            - detect
            - record
            - rtmp
    detect:
      width: 1280
      height: 720
      fps: 10
      
    objects:
      track:
        - person
        - dog
        - cat
        - bicycle
      filters:
        person:
          threshold: 0.7
        dog:
          threshold: 0.7
    record:
      enabled: True
      retain:
        days: 0
        mode: active_objects
      events:
        retain:
          default: 3
          mode: active_objects
      
    motion:
      improve_contrast: True

You could also try playing with the min_score value.

eg:

        dog:
          min_score: 0.6
          threshold: 0.85

Yeah the coral won’t improve the accuracy as the same algorithm is being run, it’s just the coral improves speed and lowers power usage as it’s purpose made hardware for the task

2 Likes

You’ll need to refer to the reducing false positives section of the frigate docs: Reducing false positives | Frigate

2 Likes

Thanks, I’m not even aware of min_score values. I’ll have to read up on it. So far all the false positives were at the 0.7 threshold:.

Got it. I’ll read up on frigate docs per your suggestions. Thanks!