Face recognition "no recent attempts" on P100 GPU + best YOLO for 9cam Frigate? (v0.17-0)

[lorenzo-torregiani]

Hi community,

Setup: Frigate 0.17-0 on Dell R730xd + Tesla P100 16GB, 9 Dahua+PTZ cameras, go2rtc RTSP restream, ONNX detector GPU.

Issue 1 - Face Recognition: Perfect person detection on cam9 (PTZ, "corridoio" zone working, debug shows stable person boxes >0.7 score), BUT Face Library always shows "No recent face recognition attempts". Zero faces extracted (known/unknown) despite frontal people at 2-5m.

Cam9 config:

text
cameras:
cam9:
ffmpeg:
hwaccel_args: preset-nvidia-h264
inputs: [rtsp://127.0.0.1:8554/cam9]
detect:
enabled: true
width: 1280
height: 720
objects:
track: [person]
filters:
person: {min_score: 0.5, threshold: 0.7}
motion:
threshold: 60
contour_area: 100
improve_contrast: true

face_recognition:
enabled: true
model_size: small # tested large
min_area: 500 # tested 50-200
recognition_threshold: 0.9
save_attempts: 200
blur_confidence_filter: true

detectors:
onnx: {type: onnx, device: GPU}
model:
model_type: yolo-generic
width: 416
height: 416
path: /models/yolo11s416.onnx # COCO-80
Tests: GPU <50% (nvidia-smi), no FaceNet/ONNX errors, api/stats face_detection=0, person events working.

Issue 2 - YOLO for P100: Current yolo11s416.onnx ~80FPS but saturates GPU, blocks face_recognition.

Questions:

Why zero faces extracted from person tracks? YOLO COCO lacks "face" class but docs say face_recognition should work standalone. Need custom model/Frigate+?

Best YOLO ONNX for P100 16GB server dell730xd?

Real P100 performance: FPS/VRAM usage on similar setups?

Logs: "cam9 track ['face'] not supported by model" (removed → same issue).

Debug screenshots (cam9+person in frame) available. Thanks!

yolo1

[inkeep-ai-assistant[bot]]

I can help clarify the face recognition requirements and YOLO model selection for your P100 setup.

Face Recognition Issue

Your face recognition setup has a critical configuration problem. When running a COCO model (like your yolo11s416.onnx) that does not include face as a detectable label, face detection runs via CV2 using a lightweight DNN model on the CPU(1). However, you should not define face in your list of objects to track when using a COCO model(1).

The key requirement: Frigate needs to first detect a person before it can detect and recognize a face(1). Your person detection appears to be working, but face detection may be failing due to several factors:

Resolution concerns: Face recognition runs on the camera stream defined by the detect role in your config(1). Your detect resolution is 1280x720, which should be sufficient, but the min_area of 500 pixels may be too high for faces at 2-5m distance(1). The default is 500 pixels, but depending on your camera's detect stream resolution, you may need to decrease this value(1).

Model requirements: When running a default COCO model or another model that does not include face as a detectable label, face detection will run via CV2 using a lightweight DNN model that runs on the CPU(1). You do not need a custom model or Frigate+ for basic face recognition - it should work with your current COCO model.

Try these adjustments:

face_recognition:
  enabled: true
  model_size: small
  min_area: 200  # Lower threshold
  detection_threshold: 0.6  # Lower from default 0.7
  recognition_threshold: 0.9

(1)

YOLO Model Selection for P100

For YOLO models with ONNX on GPU, several options are supported(2):

YOLOv9: Supported with the yolo-generic model type(2). Configuration example:

detectors:
  onnx:
    type: onnx
    device: GPU

model:
  model_type: yolo-generic
  width: 320
  height: 320
  input_tensor: nchw
  input_dtype: float
  path: /config/model_cache/yolo.onnx
  labelmap_path: /labelmap/coco-80.txt

(2)

YOLO-NAS: Also supported but may have different performance characteristics(2).

Model size consideration: Your current 416x416 model may be oversized for your needs. The documentation shows examples using 320x320 models(2), which would reduce GPU load significantly and leave headroom for face recognition processing.

The P100's 16GB VRAM should handle multiple models simultaneously. Consider using a smaller detect resolution (320x320) to reduce GPU saturation and allow the CPU-based face detection to process frames more efficiently.

Important: When using Frigate+ models, you should not define model parameters in your config except for path(2).

📚 Sources:

1. Face Recognition
2. Supported Hardware

---

Have another question? Just tag @inkeep.

[hawkeye217]

Have you uploaded at least one face? This is required. See the official docs:

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

[lorenzo-torregiani]

@hawkeye217

mqtt:
host: [MQTT_IP]
user: mqtt
password: [REMOVED]

go2rtc:
streams:
cam1:
- ffmpeg:rtsp://[CREDENTIALS_REMOVED]@[CAM_IP]/cam/realmonitor?channel=1&subtype=0
cam2:
- ffmpeg:rtsp://[CREDENTIALS_REMOVED]@[CAM_IP]/cam/realmonitor?channel=2&subtype=0
cam3:
- ffmpeg:rtsp://[CREDENTIALS_REMOVED]@[CAM_IP]/cam/realmonitor?channel=3&subtype=0
cam4:
- ffmpeg:rtsp://[CREDENTIALS_REMOVED]@[CAM_IP]/cam/realmonitor?channel=4&subtype=0
cam5:
- ffmpeg:rtsp://[CREDENTIALS_REMOVED]@[CAM_IP]/cam/realmonitor?channel=5&subtype=0
cam6:
- ffmpeg:rtsp://[CREDENTIALS_REMOVED]@[CAM_IP]/cam/realmonitor?channel=6&subtype=0
cam7:
- ffmpeg:rtsp://[CREDENTIALS_REMOVED]@[CAM_IP]/cam/realmonitor?channel=7&subtype=0
cam8:
- ffmpeg:rtsp://[CREDENTIALS_REMOVED]@[CAM_IP]/cam/realmonitor?channel=8&subtype=0
cam9:
- ffmpeg:rtsp://[CREDENTIALS_REMOVED]@[CAM_IP]:554/h264Preview_01_main

lpr:
enabled: true
min_area: 1500
recognition_threshold: 0.85
format: ^[A-Z]{1,2}[0-9]{1,3}[A-Z]{0,2}$
match_distance: 1
known_plates:
Car1:
- [PLATE1]
- [PLATE2]
- [PLATE3]
Motorcycle1:
- [PLATE1]
- [PLATE2]

cameras:
cam1:
ffmpeg:
hwaccel_args: preset-nvidia-h264
inputs:
- path: rtsp://127.0.0.1:8554/cam1
input_args: preset-rtsp-restream
detect: { enabled: false }
record: { enabled: false }

cam2:
ffmpeg:
hwaccel_args: preset-nvidia-h264
inputs:
- path: rtsp://127.0.0.1:8554/cam2
input_args: preset-rtsp-restream
detect: { enabled: true }
record: { enabled: false }
snapshots: { enabled: true }
motion:
threshold: 76
contour_area: 89
improve_contrast: true
lpr:
enabled: true
min_area: 1500

cam3-8: stesse config di cam1 (detect/record disabilitati)

cam9:
ffmpeg:
hwaccel_args: preset-nvidia-h264
inputs:
- path: rtsp://127.0.0.1:8554/cam9
input_args: preset-rtsp-restream
detect:
enabled: true
width: 1280
height: 720
face_recognition: { enabled: true }
record: { enabled: false }
snapshots: { enabled: true }
motion:
threshold: 60
contour_area: 100
improve_contrast: true
onvif:
host: [CAM_IP]
port: 8000
user: [REMOVED]
password: [REMOVED]
autotracking:
enabled: true
calibrate_on_startup: true
zooming: disabled
zoom_factor: 0.3
required_zones: [corridoio]
return_preset: home
timeout: 10
zones:
corridoio:
coordinates: 0.317,0.393,0.358,0.377,0.406,0.392,0.532,0.996,0.196,0.991,0.222,0.659
loitering_time: 0

detectors:
onnx:
type: onnx
device: GPU

model:
model_type: yolo-generic
width: 416
height: 416
input_pixel_format: rgb
input_tensor: nchw
input_dtype: float
path: /models/yolo11s416.onnx
labelmap_path: /labelmap/coco-80.txt

detect: { enabled: true }
version: 0.17-0

face_recognition: enabled: true model_size: large min_area: 50 detection_threshold: 0.4 recognition_threshold: 0.6 device: None

semantic_search:
enabled: true
model_size: large

classification:
bird: { enabled: false }

my config non working

[hawkeye217]

You've specified device: None in your face_recognition config.

You should specify device: CPU or device: GPU.

[lorenzo-torregiani]

@hawkeye217
mqtt:
host: <MQTT_HOST>
user: mqtt
password: <MQTT_PASSWORD>

go2rtc:
streams:
cam1:
- ffmpeg:rtsp://<CAM_USER>:<CAM_PASS>@<CAM_IP>/cam/realmonitor?channel=1&subtype=0
cam2:
- ffmpeg:rtsp://<CAM_USER>:<CAM_PASS>@<CAM_IP>/cam/realmonitor?channel=2&subtype=0
cam3:
- ffmpeg:rtsp://<CAM_USER>:<CAM_PASS>@<CAM_IP>/cam/realmonitor?channel=3&subtype=0
cam4:
- ffmpeg:rtsp://<CAM_USER>:<CAM_PASS>@<CAM_IP>/cam/realmonitor?channel=4&subtype=0
cam5:
- ffmpeg:rtsp://<CAM_USER>:<CAM_PASS>@<CAM_IP>/cam/realmonitor?channel=5&subtype=0
cam6:
- ffmpeg:rtsp://<CAM_USER>:<CAM_PASS>@<CAM_IP>/cam/realmonitor?channel=6&subtype=0
cam7:
- ffmpeg:rtsp://<CAM_USER>:<CAM_PASS>@<CAM_IP>/cam/realmonitor?channel=7&subtype=0
cam8:
- ffmpeg:rtsp://<CAM_USER>:<CAM_PASS>@<CAM_IP>/cam/realmonitor?channel=8&subtype=0
cam9:
- ffmpeg:rtsp://<CAM9_USER>:<CAM9_PASS>@<CAM9_IP>:554/h264Preview_01_main

lpr:
enabled: true
min_area: 1500
recognition_threshold: 0.85
format: ^[A-Z]{1,2}[0-9]{1,3}[A-Z]{0,2}$
match_distance: 1
known_plates:
Car1:
- GW503
- YW503
- CW503
Motorcycle1:
- FJ366
- FJ30

cameras:
cam1:
ffmpeg:
hwaccel_args: preset-nvidia-h264
inputs:
- path: rtsp://<CAM_USER>:<CAM_PASS>@127.0.0.1:8554/cam1
input_args: preset-rtsp-restream
detect:
enabled: false
record:
enabled: false

cam2:
ffmpeg:
hwaccel_args: preset-nvidia-h264
inputs:
- path: rtsp://127.0.0.1:8554/cam2
input_args: preset-rtsp-restream
detect:
enabled: true
record:
enabled: false
snapshots:
enabled: true
motion:
threshold: 76
contour_area: 89
improve_contrast: true
lpr:
enabled: true
min_area: 1500

cam3:
ffmpeg:
hwaccel_args: preset-nvidia-h264
inputs:
- path: rtsp://127.0.0.1:8554/cam3
input_args: preset-rtsp-restream
detect:
enabled: false
record:
enabled: false

cam4:
ffmpeg:
hwaccel_args: preset-nvidia-h264
inputs:
- path: rtsp://<CAM_USER>:<CAM_PASS>@127.0.0.1:8554/cam4
input_args: preset-rtsp-restream
detect:
enabled: false
record:
enabled: false

cam5:
ffmpeg:
hwaccel_args: preset-nvidia-h264
inputs:
- path: rtsp://<CAM_USER>:<CAM_PASS>@127.0.0.1:8554/cam5
input_args: preset-rtsp-restream
detect:
enabled: false
record:
enabled: false

cam6:
ffmpeg:
hwaccel_args: preset-nvidia-h264
inputs:
- path: rtsp://<CAM_USER>:<CAM_PASS>@127.0.0.1:8554/cam6
input_args: preset-rtsp-restream
detect:
enabled: false
record:
enabled: false

cam7:
ffmpeg:
hwaccel_args: preset-nvidia-h264
inputs:
- path: rtsp://<CAM_USER>:<CAM_PASS>@127.0.0.1:8554/cam7
input_args: preset-rtsp-restream
detect:
enabled: false
record:
enabled: false

cam8:
ffmpeg:
hwaccel_args: preset-nvidia-h264
inputs:
- path: rtsp://<CAM_USER>:<CAM_PASS>@127.0.0.1:8554/cam8
input_args: preset-rtsp-restream
detect:
enabled: false
record:
enabled: false

cam9:
ffmpeg:
hwaccel_args: preset-nvidia-h264
inputs:
- path: rtsp://<CAM_USER>:<CAM_PASS>@127.0.0.1:8554/cam9
input_args: preset-rtsp-restream
detect:
enabled: true
width: 1280
height: 720
face_recognition:
enabled: true
record:
enabled: false
snapshots:
enabled: true
motion:
threshold: 60
contour_area: 100
improve_contrast: true
onvif:
host: <CAM9_IP>
port: 8000
user: <ONVIF_USER>
password: <ONVIF_PASS>
autotracking:
enabled: true
calibrate_on_startup: true
zooming: disabled
zoom_factor: 0.3
required_zones:
- corridoio
return_preset: home
timeout: 10
zones:
corridoio:
coordinates: 0.317,0.393,0.358,0.377,0.406,0.392,0.532,0.996,0.196,0.991,0.222,0.659
loitering_time: 0

detectors:
onnx:
type: onnx
device: GPU

model:
model_type: yolo-generic
width: 416
height: 416
input_pixel_format: rgb
input_tensor: nchw
input_dtype: float
path: /models/yolo11s416.onnx
labelmap_path: /labelmap/coco-80.txt

detect:
enabled: true

version: 0.17-0

face_recognition:
enabled: true
model_size: large
min_area: 50
detection_threshold: 0.4
recognition_threshold: 0.6
device: GPU

semantic_search:
enabled: true
model_size: large

classification:
bird:
enabled: false
my config frigate kubernetes facial recognition doesn't work

[hawkeye217]

I would suggest starting with a basic face_recognition config - use the small model, and set device to CPU.

I also noticed that object detection is disabled on your cameras. In order for face recognition to run, Frigate must first detect a person. You should enable object detection.

[lorenzo-torregiani]

I've already tried using small but it still doesn't work. Also, cam9 detects people perfectly fine - detection is enabled and works very well

[hawkeye217]

Did you try device: CPU as well?