accuracy difference between pytorch and opencv

[taylan24]

Hello, Firstly, thank you very much for your work. I trained yolov3-tiny for one class with my own data and image size is 640 in training time, and the results are pretty good when I use detect.py with an image size of 640. I used the recent repository by the way. However, if I use OpenCV's dnn library, the accuracy is decreased by about 3 %. the image sizes, confidence thresholds and nms thresholds are the same. in opencv: blob = cv2.dnn.blobFromImage(img, 1.0/255.0, (640,448), (0,0,0), True, crop=False,ddepth=cv2.CV_32F) so the depth is float 32, and in detect.py the depth is torch.float32. I also used (640,640) for blob size above, nothing changed. I do not know how to solve this, and I did not find the difference in inference.

Thank you in advance.

[glenn-jocher]

@taylan24 I don't understand your question.

We do not support external functionality as you mentioned, if you have a question regarding cv2 you should raise it on the cv2 repo.

[github-actions[bot]]

This issue is stale because it has been open 30 days with no activity. Remove Stale label or comment or this will be closed in 5 days.

[sky-fly97]

你好, 首先，非常感谢你的工作。 我训练了yolov 3--用我自己的数据和图像大小来训练一个类，在训练时间里，图像大小是640，当我使用像大小为640的探测时，结果是很好的。顺便说一下，我使用了最近的存储库。 然而，如果我使用OpenCV的系统库，准确率会下降约3%。 图像大小、置信度阈值和nms阈值是相同的。 在OpenCV中： BLOB=cv2.dnn.bobFromImage(IMG，1.0/255.0，(640,448)，(0，0，0)，真，作物=假，d纵深=cv2.CV_32F) 所以深度是浮动32，在探测中，深度是torch.Float 32。 我还使用了(640,640)以上的BLOB大小，没有什么变化。 我不知道如何解决这个问题，我也没有发现推论上的区别。

提前谢谢。

Hello, I want to know how you can use dnn to read and infer, because I have an error when using dnn to read yolov3-spp.onnx, can I have a look at your opencv inference code? thank you very much!

[XiaoLaoDi]

@taylan24 would you mind to share the opencv inference code ?