yolov8Post-processing

[zjkhahah]

I printed the loaded model, but it seems that many confidence levels corresponding to categories are 1. Is this normal? Or something went wrong?
The model uses official yolov8s

##############################
ncnn::Extractor ex = yolo.create_extractor();
ex.input("in0", in_pad);
ncnn::Mat out;
ncnn::Mat transpose_out;
ex.extract("out0", out);
transpose(out, transpose_out);

for (int i = 0; i < transpose_out.h; i++) {
for (int j = 0; j < transpose_out.w; j++) {
printf("%f,", transpose_out[i *(4+ num_class) +j]);
}
printf("\n");
}
printf("width=%d,height=%d,channels=%d\n", transpose_out.w, transpose_out.h, transpose_out.c);
return 0;
}
############################
This is the output

464.000000,736.000000,512.000000,736.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,1.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.500000,0.500000,1.000000,0.500000,0.500000,0.500000,0.500000,1.000000,0.500000,1.000000,0.500000,0.500000,0.500000,0.500000,0.500000,0.500000,0.000000,0.500000,1.000000,0.500000,0.500000,0.500000,0.500000,0.500000,0.000000,0.000000,0.000000,0.000000,1.000000,0.500000,0.000000,0.500000,1.000000,0.500000,0.000000,1.000000,
496.000000,688.000000,512.000000,832.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,1.000000,1.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.000000,0.500000,0.000000,0.000000,0.000000,0.000000,0.500000,0.500000,1.000000,0.500000,0.500000,0.500000,0.500000,0.500000,0.500000,0.500000,1.000000,0.000000,0.500000,0.000000,1.000000,0.500000,0.500000,0.500000,0.500000,1.000000,0.000000,0.000000,1.000000,0.000000,1.000000,0.000000,0.500000,1.000000,0.500000,0.000000,0.000000,
width=84,height=8400,channels=1

[whyb]

I just submitted a pull request for yolov8 example #5506
I hope it can be helpful to you.

[zjkhahah]

Thanks for the reply. I seem to have a problem running the code.

I tried to print the highest confidence level for the value of out, but the output did not seem to be correct. These confidence levels felt too high.Below is the result and code I printed

ncnn::Mat out;
ex.extract("out0", out);
cv::Mat a(out.h, out.w, CV_32FC1);
memcpy((uchar*)a.data, out.data, out.w * out.h * sizeof(float));
for (int col = 0; col < a.cols; ++col) {
float maxVal = std::numeric_limits::lowest(); // 初始化为可能的最小浮点数
for (int row = 4; row < 84; ++row) { // 从第 4 行到第 83 行（索引从 0 开始）
float value = a.at(row, col);
if (value > maxVal) {
maxVal = value; // 更新最大值
}
}
std::cout << "Column " << col << " max value from row 4 to 84 is: " << maxVal << std::endl;
}

[whyb]

Thanks for the reply. I seem to have a problem running the code.

I tried to print the highest confidence level for the value of out, but the output did not seem to be correct. These confidence levels felt too high.Below is the result and code I printed

ncnn::Mat out; ex.extract("out0", out); cv::Mat a(out.h, out.w, CV_32FC1); memcpy((uchar*)a.data, out.data, out.w * out.h * sizeof(float)); for (int col = 0; col < a.cols; ++col) { float maxVal = std::numeric_limits::lowest(); // 初始化为可能的最小浮点数 for (int row = 4; row < 84; ++row) { // 从第 4 行到第 83 行（索引从 0 开始） float value = a.at(row, col); if (value > maxVal) { maxVal = value; // 更新最大值 } } std::cout << "Column " << col << " max value from row 4 to 84 is: " << maxVal << std::endl; }

我知道这个问题，这个原因不是因为我这边的原因，而是您导出的yolov8.ncnn.* 模型是仅固定支持[1,3,640,640]的模型，并不支持动态shape输入造成的。

这个错误的模型会导致输入进去的数据为了保持图像原比例缩放，并不符合模型要求的[1,3,640,640]，实际上有可能是[1,3,640,540]，所以模型内部会错位读取图像，对结果造成严重偏差。

只是要尝试效果的话，您可以直接把输入的图片用画图工具或者cv::resize() 修改成640x640的尺寸。

pnnx本身是支持动态shape输入的，您可以去参考 https://github.com/pnnx/pnnx?tab=readme-ov-file#------detailed-options

导出正确的支持动态shape输入的模型，期待您的好消息~🤣