cv::dnn::blobFromImage(image, nchw, 1 / 255.f, size, cv::Scalar(0, 0, 0), true, false, CV_32F);函数耗时

[whiteCupz]

使用后推理时间大幅提升，但是这行代码耗时最久，cv::dnn::blobFromImage(image, nchw, 1 / 255.f, size, cv::Scalar(0, 0, 0), true, false, CV_32F)。

[ZJDATY]

我看到 有别的做法 是写cuda 代码 ，全部用显卡运算。

[triple-Mu]

我看到 有别的做法 是写cuda 代码 ，全部用显卡运算。

是的，可以尝试warpaffinekernel.

[eostos]

Hello , do u have any update about this code line ? the inference is running really fast around 1ms to 2 ms but blob functions add 10 to 18 ms which affects a lot.

[triple-Mu]

Hello , do u have any update about this code line ? the inference is running really fast around 1ms to 2 ms but blob functions add 10 to 18 ms which affects a lot.

I will update a more quick cpu version the last few days.

[triple-Mu]

This pr improve preprocess hwc->nchw by opencv cv::split. Welcome to try it and give valuable feedback!

[ZJDATY]

@triple-Mu 我如果还想添加这个处理 ，大佬可以写一下吗

image -= np.array([123.675, 116.28, 103.53], np.float32)
image /= np.array([58.395, 57.12, 57.375], np.float32)

[triple-Mu]

@triple-Mu 我如果还想添加这个处理 ，大佬可以写一下吗

image -= np.array([123.675, 116.28, 103.53], np.float32)
image /= np.array([58.395, 57.12, 57.375], np.float32)

这个原版yolov8没有的吧？ 你可以尝试在 convertTo中修改alpha和beta试试