Did anyone got onnx yolov7 working with opencv?

[holger-prause]

Hello, i am wondering if anyone of you could succesfully read in yolov7 model with the opencv dnn module and got predictions with it?

The problems i am facing:

• opencv complains about missing layers after exporting to onxxx with --grid and --end2end
• when exporting without those flags i got an array back with 3 entries with the following shapes:
  <class 'tuple'>: (1, 3, 40, 40, 85)
  <class 'tuple'>: (1, 3, 20, 20, 85)
  <class 'tuple'>: (1, 3, 80, 80, 85)

I honestly dont know what to do with this result, 80 seems to correspond to the amount of classes, but in which array are the confidences / bounding boxes? 85- 80 is 5 - maybe the first 80 entries are the confidences for each class while the last 5 entries are classId, x, y, w, h?

Looking at the numbers in the array i am a bit skeptical, some are negative - shouldnt be a confidence be between 0 and 1?
Any further insight or even an opencv example using opencvs net.forward would be highly welcome.

[microboym]

#49 (comment)

seems that someone has already come up with a solution.

[holger-prause]

Thank you very much - i came across this posting - but i disregarded it as i have to switch to a special branch for this and i wanted to stick to the latest master version. Still i could port this solution to master or just use the solution supplied here - https://colab.research.google.com/github/WongKinYiu/yolov7/blob/main/tools/YOLOv7onnx.ipynb#scrollTo=07vh3pUa1ccC which just uses the onnx runtime to do the predictions (no opencv dnn module)

I am still a bit unhappy and unsure which way to got - ideally usinng opencvs dnn module on master branch - but i am a bit lost there.

[microboym]

AlexeyAB/darknet#8595 (comment)

It seems possible to convert yolov7 model into darknet format(.cfg and .weights), and maybe you can simply load the darknet model with opencv using cv2.dnn.readNetFromDarknet("yolov7.cfg", "yolov7.weights").