[Config Support]: the usual tree as a person. Is it a Coral 'problem'?

[ostracizado]

Describe the problem you are having

It's a recurrent problem I keep having with a specific tree...
A static object, aka tree, is, at certain periods of the day, identified as a person.
There no previous movement, or identifications.

The event is not created, but the "person count", "person occupancy" and entity: image.camera_person, are affected.
As automations/notifications/TTS, in Home Assistant, regarding the Frigate's Integration, are affected by those what I want to know is if there's anything I can do to avoid it.

Ex:

I did the object mask, at the bottom center of the bounding box. I tried a small and large object mask, I tried various forms (triangle, rectangle, etc.), and so on... But it seems that it does not affect the "person count", "person occupancy" and entity: image.camera_person.

Is there any solution I'm not aware? Or nothing in the config file will affect those?

Version

0.13.2

Frigate config file

...  
outside_gate:
    ffmpeg:
      inputs:
        - path: rtsp:https://127.0.0.1:8554/outside_gate #hd 1920x1080 - stream1 2k 2304x1296
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
    detect:
    #  stationary:
    #    interval: 10
      width: 1920
      height: 1080
      fps: 5
    snapshots:
      enabled: true
      timestamp: false
      bounding_box: true
      required_zones:
        - cabin_2
        - path
        - gate
      #crop: True
      retain:
        default: 3
    record:
      enabled: True
      retain:
        days: 2 
      events:
        retain:
          default: 4
          mode: active_objects 
        required_zones:
         - cabin_2
         - path
         - gate
        pre_capture: 3
        post_capture: 10 
    objects:
      track:
        - person
        - dog
        - cat
        - car
        - motorcycle
      filters:
        person:
          mask:
            #- 1563,364,1418,333,1540,136
            - 1605,246,1483,394,1459,209
            - 1670,258,1565,258,1618,396
        dog:
          mask:
            - 1454,299,1585,292,1621,0,1515,0
    motion:
      improve_contrast: true 
      mask:
        - 0,69,616,63,612,0,0,0
        - 201,408,369,252,99,74,0,57
        - 361,109,92,391,54,678,304,772,843,723,937,55
    zones:
      cabin_2:
        coordinates: 0,784,795,784,1064,487,1920,505,1920,1080,0,1080
        objects:
          - person
          - dog
          - cat
        filters:
          dog:
            min_area: 5000
          cat:
            max_area: 18000
            #min_score: 0.4
            #threshold: 0.5
          person:
            min_area: 10000 
      path:
        coordinates: 1920,498,1088,463,987,426,1106,149,1523,102,1609,114,1920,183,1920,486
        objects:
          - person
          - dog
          - cat
        filters:
          dog:
            min_area: 5000
          cat:
            max_area: 4000 
          person:
            min_score: 0.4
            max_area: 6000
      gate:
        coordinates: 1366,0,1108,0,999,131,1333,138
        objects:
          - person
          #- dog
          #- cat
          - car
          - motorcycle
        filters:
          person:
            min_score: 0.5
            min_area: 2000
            max_area: 18000
          car:
            min_score: 0.75

Relevant log output

---

Frigate stats

No response

Operating system

HassOS

Install method

HassOS Addon

Coral version

USB

Any other information that may be helpful

No response

[NickM-27]

this isn't a "coral" issue as corals just run the model they are given. The problem is that the default model is not trained on images from security cameras. This leads to false positives. Examples of the images that the default model are trained on

In this case if the object is still detected then the object mask is not being placed correctly. Can you show where the object mask is placed

[ostracizado]

At the moment I'm trying with these two:

[NickM-27]

these don't seem like the best shapes. Remember that it depends on the bottom center of the tree.

would be better with one continuous rectangle along the bottom of where the bounding box usually is

[irekzielinski]

this isn't a "coral" issue as corals just run the model they are given. The problem is that the default model is not trained on images from security cameras. This leads to false positives. Examples of the images that the default model are trained on

Would the alternative be to subscribe for frigate+ to get a better model?

[NickM-27]

Yes, that's another good way to improve detection results