Can Frigate do facial recognition on an NVIDIA 4080?

[curtiswilliams]

It's trying to run OpenVINO for face detect and that's not compatible.

[inkeep-ai-assistant[bot]]

Frigate supports facial recognition and can leverage GPU acceleration for this enrichment task, but the framework used for GPU acceleration on NVIDIA GPUs is TensorRT, not OpenVINO. OpenVINO is designed for Intel hardware and is not compatible with NVIDIA GPUs(1). For NVIDIA GPUs, you should run the Frigate Docker image tagged with -tensorrt to enable GPU acceleration for enrichments like face recognition(1).

If Frigate is attempting to use OpenVINO for face detection on an NVIDIA 4080, this setup is incorrect. You should configure Frigate to use the appropriate image and settings for NVIDIA hardware. TensorRT will be automatically detected and used for enrichments when running on supported NVIDIA GPUs(1).

For reference, here’s the relevant section:

• NVIDIA GPUs: Nvidia GPUs will automatically be detected and used for enrichments in the -tensorrt Frigate image(1).

📚 Sources:

1. Enrichments

---

Have another question? Just tag @inkeep.

[hawkeye217]

That is not correct. the -tensorrt image works just fine with Intel processors. It's the same Docker image but with the TensorRT drivers included for hardware accelerated enrichments.

[curtiswilliams]

It popped an error saying I couldn't use that image because "TensorRT detector is no longer supported on amd64 system"

[hawkeye217]

The "TensorRT detector" is referring to object detection, not enrichments.

[curtiswilliams]

Sorry, can you explain? Just trying to get face detect to quit running on my CPU, it's been at 99% utilization since I tried face detect enabled in config.

[hawkeye217]

What is your current config?

[curtiswilliams]

mqtt:
enabled: true
host: mqtt
user: mqtt
password: h8soccer

detectors:
tensorrt:
type: tensorrt
device: 0
model:
path: /config/model_cache/tensorrt/model.trt
input_tensor: nchw
input_pixel_format: rgb
width: 640
height: 640
labelmap_path: /labelmap/coco-80.txt

face_recognition:
enabled: false
model_size: large

go2rtc:
streams:
woodhouse_tree_camera_detect:
- ffmpeg:rtsp://192.168.68.146:34259/7dd2a04ae7a1ba87#video=h264#hardware&bitrate=2000k
woodhouse_tree_camera_record:
- ffmpeg:rtsp://192.168.68.146:34259/1b8f351916718278#video=h264#hardware
woodhouse_house_camera_detect:
- ffmpeg:rtsp://admin:[email protected]:554/cam/realmonitor?channel=1&subtype=0#video=h264#hardware
woodhouse_house_camera_record:
- ffmpeg:rtsp://admin:[email protected]:554/cam/realmonitor?channel=1&subtype=2#video=h264#hardware
webrtc:
candidates:
- stun:8555

cameras:
woodhouse_tree_camera:
ffmpeg:
hwaccel_args: [-hwaccel, cuda, -c:v, h264_cuvid]
input_args: preset-rtsp-restream
output_args:
record: preset-record-generic
inputs:
- path: rtsp://127.0.0.1:8554/woodhouse_tree_camera_detect
roles:
- detect
- path: rtsp://127.0.0.1:8554/woodhouse_tree_camera_record
roles:
- record
detect:
enabled: true
width: 1280
height: 720
fps: 5
motion:
enabled: true
threshold: 25
contour_area: 100
objects:
track:
- person
- car
filters:
person:
min_score: 0.5
threshold: 0.7
car:
min_score: 0.5
threshold: 0.7
record:
enabled: true
retain:
days: 7
mode: all
snapshots:
enabled: true
bounding_box: true
retain:
default: 30
woodhouse_house_camera:
ffmpeg:
hwaccel_args: [-hwaccel, cuda, -c:v, h264_cuvid]
input_args: preset-rtsp-restream
output_args:
record: preset-record-generic
inputs:
- path: rtsp://127.0.0.1:8554/woodhouse_house_camera_detect
roles:
- detect
- path: rtsp://127.0.0.1:8554/woodhouse_house_camera_record
roles:
- record
detect:
enabled: true
width: 1280
height: 720
fps: 5
motion:
enabled: true
threshold: 25
contour_area: 100
objects:
track:
- person
- car
filters:
person:
min_score: 0.5
threshold: 0.7
car:
min_score: 0.5
threshold: 0.7
record:
enabled: true
retain:
days: 7
mode: all
snapshots:
enabled: true
bounding_box: true
retain:
default: 30

database:
path: /media/frigate/database/frigate.db

detect:
enabled: true

auth:
enabled: false

logger:
default: debug

version: 0.16-0

[curtiswilliams]

Getting your yolo.onnx from any particular place?

[hawkeye217]

As I said above:

You will need to generate and download a yolov9 model according to the docs.

[curtiswilliams]

I keep running into -- 2025-09-28 15:33:52.942718990 Key : detectors -> onnx -> model -> input_type --

Got the model in place curtis@curtis-linux:~/frigate$ ls -lh ~/frigate/config/model_cache/
total 586M
-rw-r--r-- 1 root root 28M Sep 25 17:15 ba01e988083d8a3359f77f085cc0c218
-rw-r--r-- 1 root root 1.1K Sep 25 17:15 ba01e988083d8a3359f77f085cc0c218.json
-rw-r--r-- 1 curtis curtis 621 Sep 28 14:35 coco-80.txt
-rw-r--r-- 1 curtis curtis 883 Sep 28 15:16 Dockerfile
drwxr-xr-x 2 curtis curtis 4.0K Sep 25 12:20 facedet
drwxr-xr-x 3 root root 4.0K Sep 28 13:09 openvino
drwxr-xr-x 2 root root 4.0K Sep 27 15:42 tensorrt
drwxrwxr-x 6 curtis docker 4.0K Sep 28 13:32 yoloenv
-rw-rw-r-- 1 curtis docker 266M Sep 28 12:45 yolo.onnx
-rw-r--r-- 1 curtis curtis 195M Sep 28 15:19 yolov9.onnx
-rw-rw-r-- 1 curtis docker 99M Feb 18 2024 yolov9.pt

Config:
mqtt:
enabled: true
host: mqtt
user: mqtt
password: h8soccer

#################################################################

DETECTOR CONFIGURATION

#################################################################
detectors:
onnx:
type: onnx
device: 0
model:
path: /config/model_cache/yolov9.onnx
input_tensor: nchw
input_pixel_format: rgb
input_type: float32
width: 640
height: 640
labelmap_path: /config/model_cache/coco-80.txt
preprocessors:
- type: normalize
mean: [0.0, 0.0, 0.0]
std: [255.0, 255.0, 255.0]

go2rtc:
streams:
woodhouse_tree_camera_detect:
- ffmpeg:rtsp://192.168.68.146:34259/7dd2a04ae7a1ba87#video=h264#hardware=1
woodhouse_tree_camera_record:
- ffmpeg:rtsp://192.168.68.146:34259/1b8f351916718278#video=h264#hardware=1
woodhouse_house_camera_detect:
- ffmpeg:rtsp://admin:[email protected]:554/cam/realmonitor?channel=1&subtype=0
woodhouse_house_camera_record:
- ffmpeg:rtsp://admin:[email protected]:554/cam/realmonitor?channel=1&subtype=0
webrtc:
candidates:
- stun:8555

#################################################################

CAMERA CONFIGURATION

#################################################################
cameras:
woodhouse_tree_camera:
ffmpeg:
hwaccel_args: [-hwaccel, cuda, -c:v, h264_cuvid]
input_args: -fflags nobuffer -flags low_delay -rtsp_transport tcp
output_args:
record: preset-record-generic
inputs:
- path: rtsp://127.0.0.1:8554/woodhouse_tree_camera_detect
roles: [detect]
- path: rtsp://127.0.0.1:8554/woodhouse_tree_camera_record
roles: [record]
detect:
enabled: true
width: 1280
height: 720
fps: 5
motion:
enabled: true
threshold: 25
contour_area: 100
objects:
track: [person, car]
filters:
person:
min_score: 0.5
threshold: 0.7
car:
min_score: 0.5
threshold: 0.7
record:
enabled: true
retain:
days: 7
mode: all
snapshots:
enabled: true
bounding_box: true
retain:
default: 30

woodhouse_house_camera:
ffmpeg:
hwaccel_args: [-hwaccel, cuda, -c:v, h264_cuvid]
input_args: -fflags nobuffer -flags low_delay -rtsp_transport tcp
output_args:
record: preset-record-generic
inputs:
- path: rtsp://127.0.0.1:8554/woodhouse_house_camera_detect
roles: [detect]
- path: rtsp://127.0.0.1:8554/woodhouse_house_camera_record
roles: [record]
detect:
enabled: true
width: 1280
height: 720
fps: 5
motion:
enabled: true
threshold: 25
contour_area: 100
objects:
track: [person, car]
filters:
person:
min_score: 0.5
threshold: 0.7
car:
min_score: 0.5
threshold: 0.7
record:
enabled: true
retain:
days: 7
mode: all
snapshots:
enabled: true
bounding_box: true
retain:
default: 30

#################################################################

DATABASE & RETENTION

#################################################################
database:
path: /media/frigate/database/frigate.db

detect:
enabled: true

#################################################################

AUTH / LOGGING

#################################################################
auth:
enabled: false

logger:
default: debug

[NickM-27]

your config is invalid, please follow the config as it is in the docs / the example that was pasted above. You are indenting model and everything under it to be under the detector. This is wrong, you should leave that at the top level