ONNX export for single-class YOLOv7: confidences not normalized (Frigate/OpenVINO integration issue)

[kerotonic]

Hello,

I’ve trained a YOLOv7 single-class model for rodent detection (class: rat) and I’m exporting to ONNX for use in Frigate with OpenVINO.

Export command:

python3 export.py --weights best.pt \
  --grid --end2end --simplify \
  --topk-all 100 --iou-thres 0.45 --conf-thres 0.25 \
  --img-size 320 320 --max-wh 320

Issue:

• The exported ONNX runs, but confidence scores are raw logits instead of probabilities.
• In Frigate, this shows up as huge scores (e.g. score=248.0 (logs) or 20000% (Explore view)) and can crash tracking with NaNs.
• When I try --include-nms, the export succeeds but ONNX contains EfficientNMS_TRT nodes, which OpenVINO does not support.

What I tried:

• Suggested by AIs (ChatGPT, Gemini), attempted to patch export.py or Detect.forward_export to apply sigmoid() to objectness.
• Also tried modifying models/yolo.py so that objectness and class scores go through sigmoid before export.
• Still struggling to get a clean ONNX that:
  • has normalized confidences [0–1],
  • does not depend on TensorRT-specific ops.

Questions:

• Is there an officially recommended way to export YOLOv7 for single-class use cases, ensuring objectness is normalized in ONNX?
• Has anyone tested ONNX + OpenVINO pipelines (without EfficientNMS) for YOLOv7?

Thanks for any guidance.