Using dnn+onnx in opencv4.9 outs error

1

Using opencv 4.7.0 for object detection, run the following inference:
…
ReadNetFromONNX (…)
Net. setPreferableBackend (DNN_BackEND-CUDA);
Net. setPreferableTarget (DNN_TARGetCUDA);
Net. setInput (blob);
Net. forward (outs, this. unconnectedOutLayersName);
…
The returned inference result is correct, and the coordinates, width, height, and confidence of each box are all correct.

But using opencv4.9.0 for inference, each box has a confidence level, but the coordinates and width and height are both 0.0
why? help me

2

without a link to your model I don’t know how to help you.

3

It’s not a problem with the model, any onnx model can trigger this bug.

4

Wait. Somebody certainly can solve your problem without model…

5

Version 4.9 ，The inference results of the ONNX model differ between GPU inference and CPU inference. · Issue #25512 · opencv/opencv · GitHub this same question

6

It’s not the same question

The inference results of the ONNX model differ between GPU inference and CPU inference

and

Using dnn+onnx in opencv4.9 outs error

but may be it’s the same user

7

It’s the same problem.
The same model, the same code (dnn+onnx+cuda), the same image to be inspected, but changing opencv4.7 to opencv4.9, the inference result will lose the target box coordinates and width and height.

8

http://101.35.231.33/downloads/yolov8n.onnx

1 Like