Can´t get face recognition running

[espamik]

Even if 0.16 works well regarding face recognition, after I updated to 0.17.1 even after 2 weeks reading the documentation and trying I´m sorry about; I'm not able to get it up and running.

MQTT is enabled, detector is "edgetpu", I lowered filters section, I uploaded the same photos as before as samples to define a face, I tried small and large model.

Persons are detected, Face Library is empty, no pictures, nothing.

I noticed a MQTT Frigate/model_state message and as I understand, face recognition model is not loaded:

{
"jinaai/jina-clip-v1-text_model_fp16.onnx": "downloaded",
"jinaai/jina-clip-v1-tokenizer": "downloaded",
"jinaai/jina-clip-v1-vision_model_quantized.onnx": "downloaded",
"jinaai/jina-clip-v1-preprocessor_config.json": "downloaded",
"facenet-facenet.onnx": "not_downloaded",
"paddleocr-onnx-detection.onnx": "not_downloaded",
"paddleocr-onnx-classification.onnx": "downloaded",
"paddleocr-onnx-recognition.onnx": "not_downloaded",
"paddleocr-onnx-detection_v5-small.onnx": "downloaded",
"paddleocr-onnx-recognition_v4.onnx": "downloaded",
"paddleocr-onnx-ppocr_keys_v1.txt": "downloaded",
"yolov9_license_plate-yolov9-256-license-plates.onnx": "downloaded",
"facedet-facedet.onnx": "downloaded",
"facedet-landmarkdet.yaml": "downloaded",
"facedet-arcface.onnx": "downloaded"
}

Part of my config.yaml:

detect:
enabled: true
objects:
track:
- person
filters:
person:
min_score: 0.5
threshold: 0.7
face_recognition:
enabled: true
model_size: large
detection_threshold: 0.2
min_area: 100
snapshots:
enabled: true
############################### CAMERA 103 ###############################
camera_103:
ffmpeg:
inputs:
- path: rtsp://127.0.0.1:8554/camera_103_low
roles:
- detect
- path: rtsp://192.168.50.1:8554/camera_103_high
roles:
- record
live:
streams:
Main Stream: camera_103_high
Sub Stream: camera_103_low
ui:
order: 3
detect:
enabled: true
snapshots:
enabled: true
crop: false
quality: 100
bounding_box: true
mqtt:
enabled: true
lpr:
enabled: true
objects:
track:
- person
- car
- bicycle
- motorcycle
filters:
car:
mask:
0.458,0.111,0.454,0.258,0.494,0.267,0.497,0.32,0.573,0.343,0.579,0.29,0.636,0.29,0.653,0.22,0.653,0.125,0.5,0.1

  genai:
    enabled: true
    use_snapshot: true
    object_prompts:
      person: >-
        Check if the person's mouth and nose are covered by a mask, scarf, or
        cloth.
        If the mouth and nose are clearly hidden by any material, reply strictly
        with 'MASKED'.
        If mouth and nose are visible, reply 'NOT_MASKED'.
    objects:
      - person
zones:
  Entrada:
    coordinates: 
      0.002,0.994,0.006,0.238,0.298,0.188,0.305,0.299,0.787,0.402,0.899,0.496,0.993,0.789,0.995,0.988
    inertia: 2
    loitering_time: 0
    objects: person
  Calle_Montgo:
    coordinates: 
      0.002,0.006,0.994,0.056,0.995,0.271,0.794,0.396,0.327,0.296,0.31,0.27,0.059,0.328,0.056,0.142,0.004,0.1
    inertia: 2
    loitering_time: 0
    objects:
      - bicycle
      - car
      - motorcycle
      - person

What I noticed in the debug log:

2026-04-07 12:03:16.504120366 [2026-04-07 12:03:16] frigate.data_processing.real_time.face DEBUG : Running manual face detection.
2026-04-07 12:03:16.507212280 [2026-04-07 12:03:16] frigate.data_processing.real_time.face DEBUG : Detected no faces for person object.

Any ideas, why face recognition doesn't work?

Thanks for any help!

[inkeep-ai-assistant[bot]]

Based on your debug log showing "Detected no faces for person object," the face detection step is running but not finding faces in the frames. Here are the key areas to check:

1. Detect Stream Resolution

Face recognition runs on the camera stream defined by the detect role in your config. (1) You're using camera_103_low for detection - ensure this stream resolution is high enough to capture face details. The documentation notes that accuracy improves with higher quality cameras/streams, and you should consider the DORI (Detection Observation Recognition Identification) range of your camera. (1)

2. Model Download Status

Your MQTT message shows "facenet-facenet.onnx": "not_downloaded" - this is the model used for the small face recognition option. Since you're using model_size: large, the relevant models appear to be facedet-arcface.onnx (downloaded) and facedet-facedet.onnx (downloaded), which is correct. (1)

3. Minimum Face Area

You've set min_area: 100 which is very low (default is 500 pixels). (1) While this shouldn't prevent detection, faces in your low-resolution detect stream may still be too small or unclear for the face detection model to identify.

4. Face Detection Model Considerations

Since you're using an EdgeTPU detector (not a Frigate+ model that natively detects face), face detection runs via a lightweight DNN model on the CPU. (1) In this case, you should not define face in your list of objects to track. (1)

Recommendations

• Increase your detect stream resolution to capture more face detail (1)
• Check if persons are being detected at distances where faces would realistically be identifiable
• Setting your camera stream to constant bit rate (CBR) instead of variable bit rate (VBR) may lead to better image clarity (1)
• Verify the detect stream isn't overly compressed or suffering from motion blur (1)

📚 Sources:

1. Face Recognition

---

Have another question? Just tag @inkeep.

[hawkeye217]

Have you uploaded at least one face?

https://docs.frigate.video/configuration/face_recognition/#usage

[espamik]

Thank you for help. Yes, I have. Regarding stream quality, I also bought a new Reolink RLC-1212A, so detection stream is 865x512, 7 FPS, 1024 Kbps. Even in a distance of 1m Frigate doesn´t detect a face from this stream.

[hawkeye217]

I'd suggest trying the small model.

[espamik]

Thank you, I will try and let you know what happens

[espamik]

What I tried:
1.) Changing from large to small model
2.) Using record stream also as detection stream 2688x1520 2048Kbps

Changing one high-res camera stream from record to detect increase CPU abut 10%, even if have setup a RTX for ffmpeg hwaccel.

No changes, Face recognition doesnt´ work. I have downloaded the referring snapshot files. On both, snapshot and clean snapshot I can see the face clearly, even the rtsp stream on VLC. Both snapshots have had the detect stream resolution.

What I can´t understand is, that it worked before using the same server hardware and cameras running Frigate 0.16.

Also in the debug log I can see as before:
2026-04-07 18:22:55.195349246 [2026-04-07 18:22:55] frigate.events.maintainer DEBUG : Event received: tracked_object start camera_106 1775578974.845201-8585xc
2026-04-07 18:22:55.202349743 [2026-04-07 18:22:55] frigate.embeddings.maintainer DEBUG : Received update - source_type: tracked_object, camera: camera_106, data label: person
2026-04-07 18:22:55.202609492 [2026-04-07 18:22:55] frigate.embeddings.maintainer DEBUG : Processing 2 realtime processors for object 1775578974.845201-8585xc (label: person)
2026-04-07 18:22:55.202793530 [2026-04-07 18:22:55] frigate.embeddings.maintainer DEBUG : Calling process_frame on FaceRealTimeProcessor
2026-04-07 18:22:55.202991816 [2026-04-07 18:22:55] frigate.data_processing.real_time.face DEBUG : Running manual face detection.
2026-04-07 18:22:55.204260246 [2026-04-07 18:22:55] frigate.data_processing.real_time.face DEBUG : Detected no faces for person object.

Any more ideas, what I can try?

[hawkeye217]

Did you try lowering your detection_threshold?

https://docs.frigate.video/configuration/face_recognition/#advanced-configuration

[hawkeye217]

Looking back at your config, it seems like you lowered it significantly.

Is the person being detected wearing a mask? I noticed you changed your GenAI prompt to that.

[espamik]

No, the person is not wearing a mask. It´s the intention to track this for security reasons for unknown faces. I lowered detection_threshold and also min_area just to get it working. But it didn´t changes anything, no face is detected, even if a person is detected on the high res stream.

[espamik]

I also did a full reset, deleting the database and recogation photos. After new restart I uploaded the photos again defining the first face for recognition. No changes, no face is detected if a person is detected.

[hawkeye217]

It would be helpful to see screenshots of your system metrics and enrichment metrics pages.

[espamik]

Just one numbers, using the video stream for detecting I can see in Frigate "Tracked Object Details":
Person entered Entrada
Score 82%
Ratio 0.54
Area 353596px · 8.65%
The face area has more then 1000px.

So I can´t understand, what I configured wrong, that Frigate can´t find a face on that snapshot?!?

[espamik]

Thank you again for you help. The screenshots are showing the system running all cameras.
Just to let you know, I as another test I removed ffmeg hwaccel, because that's the only difference from before running 0.16, because now I have added the RTX to the server. Same as with ffmeg hwaccel, Frigate doesn't detect any face.

[hawkeye217]

I don't see anything wrong with your metrics screenshots - they show that face recognition inference is running.

For now, I'd suggest:

• Remove min_area and detection_threshold in your face_recognition config and go back to the defaults
• Move the detect stream to something that is at least 1280x720
• Confirm faces are uploaded through the Face Library (not merely by copying files from a 0.16 setup)
• Disable LPR and GenAI temporarily to ensure the enrichments pipeline is unused for anything except face recognition

[espamik]

Hi, thank you for advice. I will try again with this settings and will let you know what happens

[espamik]

I tried again with the changes you have recommended, but I'm sorry about, same as before. Face-recognition is not working. Nothing ;-(

[hawkeye217]

To rule out something with your camera or the faces on your stream, I'd suggest setting up a dummy camera with a looping video of a person with a visible face to debug this further.

I'd suggest a video like this one: https://www.pexels.com/video/a-woman-walking-up-stairs-7224318/

[espamik]

The good news: Face recognition is working well with Frigate 0.17.1! Thank you for your help, following the idea to test just with a dummy camera I´ve got it working as expected. After some more days of research and testing, the following configuration has consistently proven to be the solution for my setup:

Adding the resolution from detect stream to each camera:

detect:
  enabled: true
  width: 640
  height: 480

Follwing sections have to be included to the config.yml, even if "enabled: false" Most important, after this changes Face recognition getting started: They have to be after the camera section:

semantic_search:
face_recognition:
lpr:
classification: