[Beta Support]: Incorrect Face Detection with ArcFace (large) on RK3588 (Orange Pi 5)

[hsea]

Describe the problem you are having

I am experiencing issues with face detection and recognition when using the ArcFace (large) model​ on an RK3588 CPU (Orange Pi Plus 5).

Expected Behavior:
The model should correctly detect and extract faces when a person is present in the frame.

Actual Behavior:
When using the ArcFace (large) model, the system often fails to extract valid faces—either extracting non-face objects or missing faces entirely when a person is clearly visible.
In the same scenario, switching to the FaceNet (small) model​ works correctly, with proper face detection and extraction.

Additionally, I observed the following:
The log shows a warning:
W Query dynamic range failed. Ret code: RKNN_ERR_MODEL_INVALID. (If it is a static shape RKNN model, please ignore the above warning message.)
This warning appears even though the model runs.

Beta Version

0.17.0 Release Candidate 2

Issue Category

Object Detection / Detectors

Frigate config file

mqtt:
  enabled: false

ffmpeg:
  hwaccel_args: preset-rkmpp

detectors:
  rknn_0:
    type: rknn
    num_cores: 0
  rknn_1:
    type: rknn
    num_cores: 0
  rknn_2:
    type: rknn
    num_cores: 0

objects:
  track:
    - person
    - bird
    - cat
  filters:
    person:
      min_area: 100
      threshold: 0.7

record:
  enabled: true
  motion:
    days: 30
  alerts:
    retain:
      days: 365
      mode: motion
  detections:
    retain:
      days: 365
      mode: motion
  sync_recordings: true


snapshots:
  enabled: true
  retain:
    default: 365

motion:
  threshold: 30
  lightning_threshold: 0.6
  contour_area: 100
  improve_contrast: true

detect:
  enabled: true
  fps: 5
  min_initialized: 10
  stationary:
    interval: 50
    threshold: 50

cameras:     # No cameras defined, UI wizard should be used
  ipc544:
    enabled: true
    ffmpeg:
      inputs:
        - path: rtsp://127.0.0.1:8554/ipc544_1
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
#        - path: rtsp://127.0.0.1:8554/ipc544_2
#          input_args: preset-rtsp-restream
#          roles:
#            - detect
    live:
      streams:
        Stream 1: ipc544_1
        Stream 2: ipc544_2
    motion:
      threshold: 30
      contour_area: 200
      improve_contrast: true
      mask:
        - 0.097,0.056,0.429,0.051,0.428,0.084,0.1,0.089
        - 0.238,0.324,0.538,0.262,0.52,0,0.226,0
        - 0.228,0.712,0.245,0.928,0.213,0.925,0.217,0.972,0.262,0.972,0.287,0.995,0.254,0.663
    objects:
      track:
        - person
        - bird
        - cat
      filters:
        person:
          min_area: 3000
          threshold: 0.7
        cat:
          min_area: 1000
          threshold: 0.5
        bird:
          min_area: 400
          threshold: 0.5

 
  aipc435hp:
    enabled: true
    ffmpeg:
      output_args:
        record: preset-record-generic-audio-aac
      inputs:
        - path: rtsp://127.0.0.1:8554/aipc435hp_1
          input_args: preset-rtsp-restream
          roles:
            - record
        - path: rtsp://127.0.0.1:8554/aipc435hp_2
          input_args: preset-rtsp-restream
          roles:
            - detect
    live:
      streams:
        Stream 1: aipc435hp_1
        Stream 2: aipc435hp_2
    motion:
      threshold: 30
      contour_area: 100
      mask: 0.098,0.044,0.427,0.044,0.427,0.087,0.095,0.085
    objects:
      track:
        - person
        - bird
        - cat
      filters:
        person:
          min_area: 5000
          threshold: 0.6
        cat:
          min_area: 1000
          threshold: 0.5
        bird:
          min_area: 400
          threshold: 0.5

version: 0.17-0
go2rtc:
  streams:
    ipc544_1:
      - rtsp://admin:password@192.168.2.11:554/stream1
    ipc544_2:
      - rtsp://admin:password@192.168.2.11:554/stream2
 
    aipc435hp_1:
      - rtsp://admin:password@192.168.2.30:554/stream1
    aipc435hp_2:
      - rtsp://admin:password@192.168.2.30:554/stream2

semantic_search:
  enabled: true
  model_size: large
face_recognition:
  enabled: true
  model_size: large
lpr:
  enabled: false
classification:
  bird:
    enabled: true

Relevant Frigate log output

2026-02-20 18:06:35.156830742  [INFO] Preparing Frigate...
2026-02-20 18:06:35.763616328  [INFO] Starting Frigate...
2026-02-20 18:06:36.952758180  �[0;93m2026-02-20 18:06:36.950662658 [W:onnxruntime:Default, device_discovery.cc:211 DiscoverDevicesForPlatform] GPU device discovery failed: device_discovery.cc:91 ReadFileContents Failed to open file: "/sys/class/drm/card1/device/vendor"�[m
2026-02-20 18:06:36.952758180  [LOGGING] Last message repeated 1 times
2026-02-20 18:06:43.882313545  [2026-02-20 18:06:43] frigate.util.config            INFO    : Checking if frigate config needs migration...
2026-02-20 18:06:44.168286594  [2026-02-20 18:06:44] frigate.util.config            INFO    : frigate config does not need migration...
2026-02-20 18:06:59.857227318  [2026-02-20 18:06:59] frigate.app                    INFO    : Starting Frigate (0.17.0-4dcd296)
2026-02-20 18:07:00.201302223  [2026-02-20 18:07:00] peewee_migrate.logs            INFO    : Starting migrations
2026-02-20 18:07:00.207148627  [2026-02-20 18:07:00] peewee_migrate.logs            INFO    : There is nothing to migrate
2026-02-20 18:07:00.309966956  [2026-02-20 18:07:00] frigate.app                    INFO    : Recording process started: 918
2026-02-20 18:07:00.381726534  [2026-02-20 18:07:00] frigate.app                    INFO    : Review process started: 931
2026-02-20 18:07:00.386273325  [2026-02-20 18:07:00] frigate.app                    INFO    : go2rtc process pid: 128
2026-02-20 18:07:02.689453800  [2026-02-20 18:07:02] frigate.app                    INFO    : Embedding process started: 941
2026-02-20 18:07:02.776660149  [2026-02-20 18:07:02] frigate.detectors.plugins.rknn INFO    : You are using yolo-nas with weights from DeciAI. These weights are subject to their license and can't be used commercially. For more information, see: https://docs.deci.ai/super-gradients/latest/LICENSE.YOLONAS.html
2026-02-20 18:07:02.776660149  [LOGGING] Last message repeated 2 times
2026-02-20 18:07:02.968833141  W rknn-toolkit-lite2 version: 2.3.2
2026-02-20 18:07:02.968833141  [LOGGING] Last message repeated 1 times
2026-02-20 18:07:02.996330050  [2026-02-20 18:07:02] frigate.app                    INFO    : Output process started: 1013
2026-02-20 18:07:03.056555819  W rknn-toolkit-lite2 version: 2.3.2
2026-02-20 18:07:03.069258658  W Query dynamic range failed. Ret code: RKNN_ERR_MODEL_INVALID. (If it is a static shape RKNN model, please ignore the above warning message.)
2026-02-20 18:07:03.071174307  [2026-02-20 18:07:03] frigate.detectors.detection_runners I       : Successfully loaded RKNN model: /config/model_cache/rknn_cache/deci-fp16-yolonas_s-rk3588-v2.3.2-2.rknn
2026-02-20 18:07:03.084456099  W Query dynamic range failed. Ret code: RKNN_ERR_MODEL_INVALID. (If it is a static shape RKNN model, please ignore the above warning message.)
2026-02-20 18:07:03.089547093  [2026-02-20 18:07:03] frigate.detectors.detection_runners I       : Successfully loaded RKNN model: /config/model_cache/rknn_cache/deci-fp16-yolonas_s-rk3588-v2.3.2-2.rknn
2026-02-20 18:07:03.121965248  [2026-02-20 18:07:03] frigate.camera.maintainer      INFO    : Camera processor started for ipc544: 1093
2026-02-20 18:07:03.151590720  W Query dynamic range failed. Ret code: RKNN_ERR_MODEL_INVALID. (If it is a static shape RKNN model, please ignore the above warning message.)
2026-02-20 18:07:03.152429546  [2026-02-20 18:07:03] frigate.detectors.detection_runners I       : Successfully loaded RKNN model: /config/model_cache/rknn_cache/deci-fp16-yolonas_s-rk3588-v2.3.2-2.rknn
2026-02-20 18:07:03.213701388  [2026-02-20 18:07:03] frigate.camera.maintainer      INFO    : Capture process started for ipc544: 1163
2026-02-20 18:07:03.596066972  [2026-02-20 18:07:03] frigate.camera.maintainer      INFO    : Camera processor started for aipc435hp: 1599
2026-02-20 18:07:03.664781244  [2026-02-20 18:07:03] frigate.camera.maintainer      INFO    : Capture process started for aipc435hp: 1656
2026-02-20 18:07:15.125064188  [2026-02-20 18:07:15] frigate.util.rknn_converter    INFO    : Found existing RKNN model: /config/model_cache/jinaai/jina-clip-v1/vision_model_fp16.rknn
2026-02-20 18:07:15.249807347  W rknn-toolkit-lite2 version: 2.3.2
2026-02-20 18:07:15.654093416  W Query dynamic range failed. Ret code: RKNN_ERR_MODEL_INVALID. (If it is a static shape RKNN model, please ignore the above warning message.)
2026-02-20 18:07:15.655958148  [2026-02-20 18:07:15] frigate.detectors.detection_runners I       : Successfully loaded RKNN model: /config/model_cache/jinaai/jina-clip-v1/vision_model_fp16.rknn

Relevant go2rtc log output (if applicable)

2026-02-20 18:06:35.706427149  [INFO] Preparing new go2rtc config...
2026-02-20 18:06:36.261332215  [WARNING] Failed to convert HomeKit config to JSON, skipping cleanup
2026-02-20 18:06:36.261664420  [INFO] Starting go2rtc...
2026-02-20 18:06:36.683034430  18:06:36.682 INF go2rtc platform=linux/arm64 revision=df95ce3 version=1.9.10
2026-02-20 18:06:36.683056305  18:06:36.682 INF config path=/config/go2rtc_homekit.yml
2026-02-20 18:06:36.684154753  18:06:36.684 INF [rtsp] listen addr=:8554
2026-02-20 18:06:36.684361251  18:06:36.684 INF [api] listen addr=:1984
2026-02-20 18:06:36.684662540  18:06:36.684 INF [webrtc] listen addr=:8555
2026-02-20 18:06:45.005770388  [INFO] Starting go2rtc healthcheck service...

Install method

Docker Compose

docker-compose file or Docker CLI command

services:
  frigate:
    container_name: frigate
#    privileged: true
    restart: unless-stopped
    stop_grace_period: 30s
    image: docker.cnb.cool/frigate-cn/frigate:0.17.0-rc2-rk
    shm_size: 512m
    security_opt:
      - apparmor=unconfined
      - systempaths=unconfined
    devices:
      - /dev/dri
      - /dev/dma_heap
      - /dev/rga
      - /dev/mpp_service
    volumes:
      - /sys/:/sys/:ro
      - /etc/localtime:/etc/localtime:ro
      - /vol1/@appconf/frigate:/config
      - /vol2/frigate:/media/frigate
      - type: tmpfs 
        target: /tmp/cache
        tmpfs:
          size: 1000000000
    ports:
      - '8971:8971'
      - '8554:8554' 
      - "8555:8555/tcp"
      - "8555:8555/udp"
    environment:
      FRIGATE_RTSP_PASSWORD: "password"
      TZ: "Asia/Shanghai"
      HF_ENDPOINT: https://huggingface.mirror.frigate-cn.video
      GITHUB_ENDPOINT: https://github.mirror.frigate-cn.video
      TF_KERAS_MOBILENET_V2_WEIGHTS_URL: https://cnb.cool/frigate-cn/mirrors/storage.googleapis/-/git/raw/main/tensorflow/keras-applications/mobilenet_v2/mobilenet_v2_weights_tf_dim_ordering_tf_kernels_0.35_224_no_top.h5

Operating system

Other Linux

CPU / GPU / Hardware

Orange Pi 5 Plus 16GB (RK3588)

Screenshots

Other functions work normally

Steps to reproduce

1. Configure Frigate to use the ArcFace (large)​ model for face recognition.
2. Upload 3-5 individual photos for each of the three people; face extraction is correct.
3. Observe face detection during person presence.
4. Note incorrect extractions.
5. Switch to FaceNet (small)​ model and observe correct behavior.

Any other information that may be helpful

There was also an installation issue with the ArcFace model:
frigate.util.rknn_converter Error during RKNN conversion: No module named 'pkg_resources'
This was resolved by manually installing setuptools inside the Docker container, but the face detection issues persist.

uname -a
Linux OPI 6.12.41-trim #227 SMP PREEMPT Thu Feb 12 08:05:38 UTC 2026 aarch64 GNU/Linux

The issue seems specific to the ArcFace model on RK3588 hardware.
Thank you for this great project

[NickM-27]

That warning is not important as it is a static shape model.

It's always possible that the model is just getting it wrong for no particular reason. It's also possible that the quantization processes reduces accuracy.

In my testing it did generally work correctly

[hsea]

Could this be related to compatibility issues with the RK3588 hardware?

I have tried multiple cameras and provided 3 photos for each of 3 individuals during training, all taken from high-resolution photos at different times using a phone.
When using the FaceNet model, recognition works normally: the same person is correctly identified, while different individuals show a low similarity rate (<50%). However, when using the ArcFace model, it often gives a very high similarity rate (99%-100%) for clearly different people—or even for images that are not human faces at all—which clearly seems abnormal.

Besides, I did not encounter this issue when using an Intel N100 CPU two month ago.

[NickM-27]

It has to convert the model so that it can run it, so the format may perform differently. I'd suggest running the small model and we can see if others report the same