[Question]: Review screen

[xavan83]

What is your question:

First of all, great thank you for your work (starting with the 0.13 version and now testing the 0.14 beta).
I look at the beta documentation of the new review screen and I don't find a way to understand why some people crossing the street in front of the camera are not detected.
I only record on event however I can see people in low resolution when the timeline is displaying at fast speed (8x). I imagine you keep some snapshots in memory to display that.
I imagine it is a good idea to switch the recording to continuous to debug that?
After that can I replay the footage with the motion/detection boxes to understand why the people are not detected? May be cross the footage with skipped frames statistics to know if the server was not under to much load at that point?

[NickM-27]

You need to include your config otherwise there's no way to help.

[xavan83]

motion:
  mqtt_off_delay: 30 
birdseye:
  enabled: true
  mode: continuous
ffmpeg:
  hwaccel_args: preset-vaapi
  output_args:
    record: preset-record-generic-audio-aac
database:
  path: /db/frigate.db
mqtt:
  enabled: true
  host: 192.168.0.16
  port: 1884
  topic_prefix: frigate
  client_id: frigate
logger:
  default: info
  logs:
    frigate.comms.dispatcher: error
    #frigate.object_detection: debug
    #frigate.record.maintainer: debug
detectors:
  coral1:
    type: edgetpu
    device: pci:1
  coral2:
    type: edgetpu
    device: pci:0
model:
  path: plus:https://xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

snapshots:
  enabled: true
  clean_copy: true
  bounding_box: false
  crop: false
  retain:
    default: 10
    objects:
      person: 30
go2rtc:
  streams:
    sonnette:
      - rtsp:https://admin:{FRIGATE_RTSP_PASSWORD}@192.168.0.70:554/Streaming/Channels/101#timeout=20
      - ffmpeg:sonnette#audio=aac#timeout=20#input=-avoid_negative_ts make_zero -fflags
        +genpts+discardcorrupt -stimeout "5000000" -use_wallclock_as_timestamps "1"
cameras:
  sonnette:
    ffmpeg:
      output_args:
        record: preset-record-generic-audio-aac
      inputs:
        - path: rtsp:https://127.0.0.1:8554/sonnette?video=copy&audio=aac
          roles:
            - detect,record
    detect:
      fps: 5
      enabled: true
      width: 1280
      height: 720
    motion:
      mask: 0.491,0.658,0.993,0.589,1,0,0,0,0,0.629
    mqtt:
      enabled: true
      timestamp: true
      bounding_box: true
      crop: false
      quality: 100
      required_zones: []
    zones:
      streetpeople:
        objects:
          - person
        inertia: 90
        coordinates: 0,720,1280,720,1280,434,0,451
      streetcar:
        objects:
          - car
        inertia: 60
        coordinates: 0,720,1280,720,1280,434,0,451
      boiteauxlettres:
        objects:
          - person
          - face
        coordinates: 1280,712,1280,720,0,720,0,712
    objects:
      filters:
        person:
          min_area: 25000
          max_ratio: 1.0
          mask:
            - 638,463,1280,422,1280,0,0,0,0,458
        car:
          min_score: 0.5
          #min_area: 100000
          mask:
            - 638,463,1280,422,1280,0,0,0,0,458
        cat:
          mask:
            - 638,463,1280,422,1280,0,0,0,0,458
        dog:
          mask:
            - 638,463,1280,422,1280,0,0,0,0,458
      track:
        - face
        - person
        - bicycle
        - car
        - motorcycle
        - bus
        - cat
        - dog
        - horse
objects:
  filters:
    person:
      min_score: 0.70
    car:
      min_score: 0.70
    dog:
      min_area: 10000
    cat:
      min_area: 10000
  track:
    - person
    - licence_plate
record:
  enabled: true
  events:
    pre_capture: 15
    post_capture: 15
    objects:
      - person
      - bicycle
      - car
      - motorcycle
      - bus
      - train
      - cat
      - dog
      - horse
    retain:
      default: 30
      mode: motion

[NickM-27]

no changes have been made to how motion detection works. And no, frigate does not limit how many objects are detected at once

[xavan83]

Does the motion mask is incorrectly used in this case?
Can I define a more sensitive are (threshold, contour_area) in the first centimeters between the road and the trees?

[NickM-27]

In general I would suggest reducing the size of the motion mask but can't make specific recommendations without seeing the camera and potentially an example video of a person being missed.

Could also be that your min_score is too high

[xavan83]

The configuration is the one above.
Can I send you in private an example video (How can I download the video file of an event in the new gui?)

Global level
objects:
filters:
person:
min_score: 0.70

motion at the camera level:
motion:
mask:
- 638,463,1280,422,1280,0,0,0,0,458 # (same for person filter)

filters at the camera level:
person:
min_area: 25000
max_ratio: 1.0
mask:
- 638,463,1280,422,1280,0,0,0,0,458

[NickM-27]

I would suggest just doing a screen recording of the debug view.

If you are having trouble with detections at all you should remove your filters

[xavan83]

I was able to download the video using export function then "Select from timeline".
I removed the filters and I will make some tests tomorrow.
Do you want to see the video?

[NickM-27]

I use the same config for v13 and v14.

in general that is not possible since many parts of the config have changed between 0.13 and 0.14

if there were errors then they would be shown in the logs

[xavan83]

I had to remove use_experimental parameter, coordinates are converted to percentage
Pretty much the same configuration
And no errors in the logs

[NickM-27]

then a current copy of the 0.14 config and screenshots showing the difference between the two screens will be necessary

[xavan83]

I don't want to share screenshots of my events (for privacy reasons of course). Is there a way to communicate these privately?

[NickM-27]

Can reach out on discord or email [email protected]