[Detector Support]: Facial recognition with distant face on 4k camera

[USER19464727]

Describe the problem you are having

I have Frigate installed on HAOS. I got facial recognition working via double-take and codeproject.ai. However, one of my cameras is a 4k wide view camera. Faces are identifiable on the full-resolution image, but despite my best efforts I cannot seem to get a true full-res image to codeproject.ai. I have tried turning off the "crop" option for snapshots and mqtt and set the height to the full height of the camera (2160). I see in the double-take logs that an image comes in but the face wasn't recognized and I can tell it's because it's too blurry since it's a small ROI on a lower-resolution image. Can someone give me some guidance?

Frigate config:

snapshots:
  enabled: true
  clean_copy: true
  timestamp: false
  bounding_box: false
  crop: false
  height: 2160
cameras:
  cam2:
    mqtt:
      crop: false
      height: 2160

Double-take config:

frigate:
  attempts:
    latest: 10
    snapshot: 10
    mqtt: true
  image:
    height: 2160

Home assistant config:

image_processing:
  - platform: codeproject_ai_object
      scale: 1.0
      roi_x_min: 0
      roi_y_min: 0
      roi_x_max: 1
      roi_y_max: 1
      crop_to_roi: False

Version

Latest for each as of 11/1/2024

Frigate config file

snapshots:
  enabled: true
  clean_copy: true
  timestamp: false
  bounding_box: false
  crop: false
  height: 2160

docker-compose file or Docker CLI command

N/A

Relevant Frigate log output

N/A

Install method

HassOS Addon

Object Detector

Coral

Screenshots of the Frigate UI's System metrics pages

Not important for this

Any other information that may be helpful

No response

[NickM-27]

the api returns the image used for detect, so if you are using a sub stream then this will be used. You will want to setup doubletake to use a full res snapshot url separately

[USER19464727]

Does this mean I would set double-take to use a particular camera stream and then configure it to take an image from that stream using an external trigger (automation)?

[NickM-27]

you setup go2rtc to read snapshots from the cameras onvif url, then use go2rtcs still frame api. Trying to integrate this into doubletake is quite difficult

[USER19464727]

If there are better alternatives for adding facial recognition to Frigate/HAOS, I'm all ears! I want to recognize faces & cars in case there's an unknown face or car in my driveway. It looked like Deepstack is not supported any more. Maybe Doods?

[NickM-27]

you could do something similar to #8165

Frigate 0.16 will also support face recognition built in

[USER19464727]

That link looks interesting, although also complicated. I like hearing about Frigate 0.16! Maybe I can be patient. Thanks.