[Detector Support]: OpenVINO bundled SSD model (ssdlite_mobilenet_v2.xml) missing labelmap causes all detections offset by 1 class — persons detected as "bicycle"

[Oasis-Enterprise]

Describe the problem you are having

When using the bundled OpenVINO SSD model with model_type: ssd, all object detections are offset by one class. A person standing directly in front of the camera is detected as "bicycle" at 91% confidence. This is because the Intel Open Model Zoo SSD MobileNet V2 model uses class 0 for background and class 1 for person, while Frigate's default internal labelmap starts at class 0 = person. No labelmap file ships alongside the bundled model and no documentation mentions this offset.

Version

0.17.1

Frigate config file

detectors:
  ov:
    type: openvino
    device: AUTO

model:
  path: /openvino-model/ssdlite_mobilenet_v2.xml
  model_type: ssd
  width: 300
  height: 300
  input_tensor: nhwc
  input_pixel_format: bgr

docker-compose file or Docker CLI command

N/A — installed via Proxmox LXC community script, not Docker. Frigate runs as a native systemd service at /etc/systemd/system/frigate.service.

Relevant Frigate log output

No errors logged. Detector starts successfully, inference speed ~9ms, detection_fps active. Silent failure — detections occur but all class IDs are offset by 1.

Install method

Proxmox via TTeck Script

Object Detector

OpenVino

Any other information that may be helpful

the fix was cat > /openvino-model/coco_labels.txt << 'EOF'
background
person
bicycle
...
EOF and then model:
path: /openvino-model/ssdlite_mobilenet_v2.xml
model_type: ssd
labelmap_path: /openvino-model/coco_labels.txt
width: 300
height: 300
input_tensor: nhwc
input_pixel_format: bgr.

[inkeep-ai-assistant[bot]]

---

Your diagnosis and fix are correct. The bundled OpenVINO SSD MobileNet V2 model uses a labelmap where class 0 is "background" and class 1 is "person," following the COCO 91-class format with background. (1)

The documentation shows the correct configuration for the default OpenVINO model:

detectors:
  ov:
    type: openvino
    device: GPU # Or NPU

model:
  width: 300
  height: 300
  input_tensor: nhwc
  input_pixel_format: bgr
  path: /openvino-model/ssdlite_mobilenet_v2.xml
  labelmap_path: /openvino-model/coco_91cl_bkgr.txt

(1)

The key is specifying labelmap_path: /openvino-model/coco_91cl_bkgr.txt — this file is included in the container and contains the 91-class COCO labelmap with background as class 0. (1)

Without explicitly setting labelmap_path, Frigate uses its default internal labelmap which doesn't account for the background class offset, causing all detections to be shifted by one class (person → bicycle, bicycle → car, etc.).

Your manual fix creating a labelmap with "background" as the first entry achieves the same result. The simpler solution is to use the bundled coco_91cl_bkgr.txt file that ships with the OpenVINO model. (1)

📚 Sources:

1. Supported Hardware

---

Have another question? Just tag @inkeep.

[NickM-27]

What you're saying is incorrect. The Frigate image does include this labelmap and the documentation specifically covers this case

https://docs.frigate.video/configuration/object_detectors#ssdlite-mobilenet-v2

Since you're using the tteck script in proxmox, which is not maintained by us, it's possible that they forgot to include this.