Openvino Custom YOLO Inference Issue

[SANTHAKUMAR91]

Hello, 

I have trained a custom darknet yolo for character recognition and successfully converted darket net model -> keras -> tensorflow -> IR <Openvino format>. I was able to do inference in openvino Yolov3 Async inference code with few custom changes on parsing yolo output. The results are same as original model. But when tried to replicated the same in c++, the results are wrong. I did small work around on the parsing output results.  I managed to get the results but they are not the same as Openvino python inference.  

Issues:

1. Predictions are deviated by little offset.

2. Certain bbox predictions are missing.

This issues are not seen in Python inference.

The Yolo Network Details: 

num = 2
coords = 4
classes = 35

anchors=  { 3.638, 5.409, 3.281, 4.764 }

Input 1 x 3 x 80 x 240 - NCHW

output = 10 x 30 x 80 - Blob Output

Kindly help me resolving the issue. Thanks  

[IntelSupport]

Hello Santhakumar91,

Thanks for reaching out. We are investigating this issue and will update you with the information soon. Meanwhile, could you share your model and inference code for us to test it on our end?

Regards,

Aznie

[SANTHAKUMAR91]

Hello, 

Thanks for the response. I have attached the model and code. 

I have used Openvino 2021.3.0 version.

[IntelSupport]

Hello Santhakumar91,

Could you please share the source of the OCR model that you used. Thanks.

Regards,

Aznie

[SANTHAKUMAR91]

Hello,

Its a customized yolo v3 using the darknet source. Trained on custom dataset.

[IntelSupport · Accepted solution]

Hi Santhakumar,

Thanks for your patience. We have tested your inference code, however, we are unable to replicate the same result as yours on the C++ inference code. You might need to change your customer code. Meanwhile, I noticed that your yolodetector.cpp is similar to our main.cpp file in Multi-Channel Object Detection Yolov3 C++ Demo. I would advise you to use it as a reference for your customer inference code as it has been tested and validated for yolo-v3 model. Also, on why both platforms give different results while running the same model- it may due to some library and component availabilities on Phyton are well covered compared to C++.

Regards,

Aznie

[IntelSupport]

Hi Santhakumar,

This thread will no longer be monitored since we have provided a solution. If you need any additional information from Intel, please submit a new question.

Regards,

Aznie

[Aznie_Intel]

Hi Santhakumar,

The developer team is looking into your Yolov3 Async C++ inference performance issues. I will update you once we obtain feedback from them.

Regards,

Aznie

[IntelSupport]

Hi Santhakumar,

Did you train the model with yolo3? We found that the C++ code is based on yolo5 and the python code s based on yolo3.

Regards,

Aznie

[SANTHAKUMAR91]

Hi Aznie,

Yes the model is trained on yolo v3 architecture. 

[IntelSupport]

Hi Santhakumar,

Could you provide us the compilation step for C++ compilation or CMAKE package so that we can directly compile?

Regards,

Aznie

[IntelSupport]

Hi Santhakumar,

Is it possible for you to share with us your Makefile or compilation step? We need it to determine the issue that happens with your C++ custom code.

Regards,

Aznie

[SANTHAKUMAR91]

Hi , This is Visual Studio Project, i have used "security_barrier_camera_demo"  project from openvino for setup and execution. Thanks 

[IntelSupport]

Hi Santhakumar,

Thank you for your reply. Do you still need any help regarding this?

Regards,

Aznie

[IntelSupport]

Hi Santhakumar,

Thank you for your question. If you need any additional information from Intel, please submit a new question as this thread is no longer being monitored.

Regards,

Aznie