[Config Support]: Perfect person detection on low res but not on higher res

[geyles]

Describe the problem you are having

Trying to increase detection on higher resolution for the purpose of better face detection and LPR. On the standard low res it detection a person fine. If I increase to a higher res it hardly ever detection a person, I've tried playing around with the min and max area settings person but this doesn't seem to have any effect. Not sure if I'm missing something really simple but I've been trying to do this on and off for weeks.

Version

16.1

Frigate config file

logger:
  default: error

#genai:
#  enabled: true
#  provider: gemini
#  api_key: ""
#  model: gemini-1.5-flash

mqtt:
  host: 192.168.1.9
  port: 1883
  topic_prefix: frigate
  client_id: frigate
  user: thatch
  password: ''
  stats_interval: 60

# Optional: Detectors configuration. Defaults to a single CPU detector
detectors:
  coral:
    type: edgetpu
    device: usb

birdseye:
  enabled: true
  restream: true
  width: 1280
  height: 720
  quality: 8
  mode: objects
  layout:
    scaling_factor: 1

ffmpeg:
#  hwaccel_args: preset-intel-qsv-h264
  hwaccel_args: preset-vaapi

detect:
  stationary:
    interval: 10
  width: 1280
  height: 720
  fps: 5
  enabled: true
  max_disappeared: 25

objects:
  track:
    - person
    - car
    - dog
  filters:
    person:
      min_area: 5000
      max_area: 100000
      min_score: 0.5
      threshold: 0.7

motion:
  threshold: 25
  contour_area: 100
  delta_alpha: 0.2
  frame_alpha: 0.2
  frame_height: 180

record:
  enabled: true
  retain:
    days: 0
  alerts:
    retain:
      days: 5
  detections:
    retain:
      days: 5
snapshots:
  enabled: true
  timestamp: false
  bounding_box: false
  crop: false
  height: 600
  required_zones: []
  retain:
    default: 10
    objects:
      person: 7
      car: 7
      dog: 7

live:
  height: 720
  quality: 8

go2rtc:
  streams:
    frontdoor_rtsphd:
      - rtsp://admin:[email protected]:554/cam/realmonitor?channel=1&subtype=0&unicast=true&proto=Onvif
#      - rtsp://admin:[email protected]:554/cam/realmonitor?channel=1&subtype=0#backchannel=0
#      - ffmpeg:frontdoorhd_rtsp#audio=opus
      - ffmpeg:frontdoor_rtsphd#audio=copy
    frontdoor_rtsp:
      - rtsp://admin:[email protected]:554/cam/realmonitor?channel=1&subtype=1&unicast=true&proto=Onvif
#      - rtsp://admin:[email protected]:554/cam/realmonitor?channel=1&subtype=1#backchannel=0      
#      - ffmpeg:frontdoor_rtsp#audio=opus
      - ffmpeg:frontdoor_rtsp#audio=copy
  webrtc:
    listen: :8555
    candidates:
      - 192.168.1.66:8555
      - stun:8555

cameras:
  frontdoor:
    birdseye:
      mode: continuous
#    genai: # <- disable GenAI for your indoor camera
#      enabled: true
#      prompt: "Analyze the {label} in these images from the {camera} security camera at the front door. Focus on the actions and potential intent of the {label}."
#      object_prompts:
#        person: "Examine the person in these images. What are they doing, and how might their actions suggest their purpose (e.g., delivering something, approaching, leaving)? If they are carrying or interacting with a package, include details about its source or destination."
#        dog: "Observe the dog in these images. Focus on its movement and intent (e.g., wandering, hunting, interacting with objects). If the dog is near the grass or engaging in any specific actions, mention it."
#      objects:
#        - person
#        - dog
    audio:
      enabled: true
      listen:
        - bark
        - fire_alarm
        - smoke_detector
        - scream
        - speech
        - yell

    ffmpeg:
      input_args: -avoid_negative_ts make_zero -fflags +genpts+discardcorrupt 
        -rtsp_transport tcp -timeout 10000000 -use_wallclock_as_timestamps 1

      inputs:
        - path: rtsp://192.168.1.66:8554/frontdoor_rtsp?video=all&audio=all
          input_args: preset-rtsp-restream
          roles:
            - detect
            - audio
            - record
#        - path: rtsp://192.168.1.66:8554/frontdoor_rtsp?video=all&audio=all
#          input_args: preset-rtsp-restream
#          roles:
#            - record
      output_args:
        #record: -f segment -segment_time 10 -segment_format mp4 -reset_timestamps 1 -strftime 1 -c copy
        record: preset-record-generic-audio-copy
    live:
      streams:
        Main stream: frontdoor_rtsp
        Sub stream: frontdoor_rtsphd
    detect:
      width: 640
      height: 360
      fps: 15
#    detect:
#      width: 1920
#      height: 1080
#      fps: 10
    best_image_timeout: 60

    zones:
      front_porch:
        coordinates: 
          0.903,0.889,0.916,0.808,0.724,0.801,0.636,0.198,0.198,0.194,0.032,0.502,0.177,1,0.023,1,0.938,1,0.959,0.962
        objects:
          - person
          - dog
        filters:
          person:
            min_area: 5000
            max_area: 100000
#            min_area: 5000
#            max_area: 1000000
            threshold: 0.7
        inertia: 3
        loitering_time: 0
      front_road:
        coordinates: 0.219,0,0.382,0,0.931,0,0.895,0.158,0.078,0.119,0.078,0
        objects:
          - person
          - car
          - dog
        filters:
          person:
            min_area: 5000
            max_area: 100000
#            min_area: 10000
#            max_area: 1000000
            threshold: 0.7
        inertia: 3
        loitering_time: 0
    mqtt:
      enabled: true
      timestamp: true
      bounding_box: true
      crop: true
      height: 540
      quality: 70
    review:
      alerts:
        required_zones:
          - front_porch
          - front_road

Relevant Frigate log output

Not necessary

Relevant go2rtc log output

Not necessary

Frigate stats

Not necessary

Operating system

Proxmox

Install method

Docker Compose

docker-compose file or Docker CLI command

Not necessary

Object Detector

Coral

Screenshots of the Frigate UI's System metrics pages

Not necessary

Any other information that may be helpful

No response

[inkeep-ai-assistant[bot]]

Based on your configuration, the issue is that your min_area and max_area filters are based on the resolution of your detect stream, not the camera's full resolution(1)(2). You're using a detect resolution of 640x360 (230,400 total pixels), but your area filters are set for a much higher resolution(1).

When you increase the detect resolution, the same physical object will have a larger pixel area in the bounding box(2). Your current filters are:

person:
  min_area: 5000
  max_area: 100000

(3)

For a 640x360 detect stream, these values work well, but when you switch to a higher resolution like 1920x1080, a person that previously had a bounding box of 50,000 pixels might now have 450,000 pixels, exceeding your max_area limit(1)(2).

To fix this for higher resolution detection:

1. Increase your area filters proportionally - if you're going from 640x360 to 1920x1080, that's roughly 9x more pixels, so multiply your area values accordingly(1):

person:
  min_area: 45000    # 5000 * 9
  max_area: 900000   # 100000 * 9

(3)

2. Use the debug view to see the actual area values of detected objects when using higher resolution(2). This will help you tune the values precisely(1).

3. Consider that motion clarity is more important than extra pixels for detection accuracy(4). The face recognition documentation notes that over-compression and motion blur are more detrimental to results than resolution(4).

The area filters can be specified as integers for width×height in pixels or as decimals representing the percentage of the frame(1)(3). Using percentages might be easier when switching between resolutions:

person:
  min_area: 0.02     # 2% of frame
  max_area: 0.43     # 43% of frame

(3)

📚 Sources:

1. Filters
2. [Config Support]: Please help with max_area for camera detection of a person
3. Full Reference Config
4. Face Recognition

---

Have another question? Just tag @inkeep.

[geyles]

I've tried but it just doesn't seem to detect when I change to higher resolution

[NickM-27]

what is your config now?

[geyles]

Same as above but changed
min_area: 0.02
max_area: 0.43
To the percentages

[NickM-27]

if you're having issues it is best to remove all filters, only add them in after detection is working