Opencv4.5.4 Inference result error

[XiangMichael]

System information (version)

When testing yolov3, when the inference width and height values are set differently, the inference result is wrong. For example, when inpWidth=480, inpHeight=320, the inference result is wrong; when inpWidth=320, inpHeight=320, the inference result is correct

[Iffa-Intel]

This is actually more related to your native YOLOv3 model instead of OpenVINO.
In order to overcome this issue, you'll need to train your data for non-square images.
This discussion thread might be useful to you: ultralytics/yolov3#126
and here is the tutorial for YOLOv3 non-square images: ultralytics/yolov3#232

If you are looking for a solution within OpenVINO, you may try to use the Reshape Inference Feature to change your model's input shape. You may refer to this Reshape Inference Feature documentation.

[alalek]

OpenCV ticket is here: opencv/opencv#19781

[XiangMichael]

OpenCV ticket is here: opencv/opencv#19781
When I did it according to opencv/opencv#19781, when "backendId = CV ::DNN ::DNN_BACKEND_INFERENCE_ENGINE, TargetId = CV ::DNN ::DNN_TARGET_CPU "Running result is normal, but" backendId = CV ::DNN ::DNN_BACKEND_INFERENCE_ENGINE, TargetId = CV ::DNN ::DNN_TARGET_OPENCL_FP16 "will crash directly

OpenCV: terminate handler is called! The last OpenCV error is:
OpenCV(4.5.4) Error: Unspecified error (Failed to initialize Inference Engine backend (device = GPU): Ngraph operation Transpose with name permute_16 has dynamic output shape on 0 port, but CPU plug-in supports only static shape) in cv::dnn::InfEngineNgraphNet::initPlugin, file D:\code\vs2017_major\opencv454_openvino\modules\dnn\src\ie_ngraph.cpp, line 742

[XiangMichael]

This is actually more related to your native YOLOv3 model instead of OpenVINO. In order to overcome this issue, you'll need to train your data for non-square images. This discussion thread might be useful to you: ultralytics/yolov3#126 and here is the tutorial for YOLOv3 non-square images: ultralytics/yolov3#232

If you are looking for a solution within OpenVINO, you may try to use the Reshape Inference Feature to change your model's input shape. You may refer to this Reshape Inference Feature documentation.

When I did it according to opencv/opencv#19781, when "backendId = CV ::DNN ::DNN_BACKEND_INFERENCE_ENGINE, TargetId = CV ::DNN ::DNN_TARGET_CPU "Running result is normal, but" backendId = CV ::DNN ::DNN_BACKEND_INFERENCE_ENGINE, TargetId = CV ::DNN ::DNN_TARGET_OPENCL_FP16 "will crash directly
OpenCV: terminate handler is called! The last OpenCV error is:
OpenCV(4.5.4) Error: Unspecified error (Failed to initialize Inference Engine backend (device = GPU): Ngraph operation Transpose with name permute_16 has dynamic output shape on 0 port, but CPU plug-in supports only static shape) in cv::dnn::InfEngineNgraphNet::initPlugin, file D:\code\vs2017_major\opencv454_openvino\modules\dnn\src\ie_ngraph.cpp, line 742

[jgespino]

Hi @XiangMichael

Apologies for the delay, is this something you are still working on? If so, please share your model and sample code to reproduce the issue.

Regards,
Jesus