search

DanielSarmiento04 / yolov10cpp

Implementation of yolo v10 in c++ std 17 over opencv and onnxruntime
73 stars 9 forks source link

Confidence level exception #4

Open xufabs opened 3 months ago

xufabs commented 3 months ago

First of all thank you for your contribution.

But there was a problem.

There are some pictures using this program for ONNXRUNTIME C++ inference, the confidence performance is too different from the yolov10 official inference.

The woman in white has only 0.5 confidence, while python has 0.8.

I've also used some other people's images, which have a 0.8 confidence with python, but only 0.1 with this project, and will even be recognized as birds.

I want to know what caused this.

onnxruntime-linux-x64-1.17.1

1

result yolov10cpp + yolov10n.onnx 👆

1 yolov10py + yolov10n.pt 👆

xufabs commented 3 months ago

First of all thank you for your contribution.

But there was a problem.

There are some pictures using this program for ONNXRUNTIME C++ inference, the confidence performance is too different from the yolov10 official inference.

The woman in white has only 0.5 confidence, while python has 0.8.

I've also used some other people's images, which have a 0.8 confidence with python, but only 0.1 with this project, and will even be recognized as birds.

I want to know what caused this.

onnxruntime-linux-x64-1.17.1

1

result yolov10cpp + yolov10n.onnx 👆

1 yolov10py + yolov10n.pt 👆

ubuntu 24.04 opencv 4.5.3

xufabs commented 3 months ago

https://github.com/UNeedCryDear/yolov8-opencv-onnxruntime-cpp I hope this can help you to debug

图片 yolov8 onnxruntime cpu cpp + yolov8n.onnx

DanielSarmiento04 commented 2 months ago

Sure, I replicate this result, and yeah, there is a different between the both

sdreamforchen commented 1 month ago

Sure, I replicate this result, and yeah, there is a different between the both

Can you directly supply ONNX model of yolov10 for downloading in your github page?

xufabs commented 1 month ago

https://github.com/Ultraopxt/yolov10cpp

maybe this can help

Sure, I replicate this result, and yeah, there is a different between the both

Can you directly supply ONNX model of yolov10 for downloading in your github page?

Sure, I replicate this result, and yeah, there is a different between the both

DanielSarmiento04 commented 2 days ago

I've some experiment with change a lot parameters and the best way at this moment I recommend you that lower the threshold confidences result

DanielSarmiento04 commented 2 days ago

https://github.com/Ultraopxt/yolov10cpp，maybe this can help

Sure, I replicate this result, and yeah, there is a different between the both

Can you directly supply ONNX model of yolov10 for downloading in your github page?

Sure, I replicate this result, and yeah, there is a different between the botbut I think that is no necessary apply a letter box to ma

It's a good idea

DanielSarmiento04 commented 2 days ago

result

With letter box implentations