yolo11和yolo8后处理方法是完全一样的。
首先安装ultralytics，然后执行下面代码导出yolo11l.onnx模型。
from ultralytics import YOLO model = YOLO('yolo11l.pt') success = model.export(format = 'onnx', opset = 11)
模型输入维度是[1,3,640,640]，输出维度是[1,84,8400]。
输出84中前4个值代表目标[x,y,w,h]，余下80个代表目标类别可能性。8400代表候选目标个数。
C++后处理代码如下，代码没有恢复比例，没有映射ID，这两步应该比较容易：
#include <iostream> #include <vector> #include <onnxruntime_cxx_api.h> #include "opencv2/opencv.hpp" using namespace std; struct CDetectBox { int id; float score; float x1; float y1; float x2; float y2; }; float compute_iou(const CDetectBox& box1, const CDetectBox& box2) { float xmin = std::max(box1.x1, box2.x1); float ymin = std::max(box1.y1, box2.y1); float xmax = std::min(box1.x2, box2.x2); float ymax = std::min(box1.y2, box2.y2); float intersection_area = std::fmax(0.0, xmax - xmin) * std::fmax(0.0, ymax - ymin); float boxArea1 = (box1.x2 - box1.x1) * (box1.y2 - box1.y1); float boxArea2 = (box2.x2 - box2.x1) * (box2.y2 - box2.y1); float union_area = boxArea1 + boxArea2 - intersection_area + 1e-10; return intersection_area / union_area; } int main() { int imgH = 640; int imgW = 640; cv::Mat image = cv::imread("test2.png", 1); cv::resize(image, image, cv::Size(imgW, imgH)); vector<float> imgData; imgData.resize(3 * imgW * imgH); for (int k = 0; k < 3; k++) { for (int i = 0; i < imgH; i++) { for (int j = 0; j < imgW; j++) { imgData[k * imgW * imgH + i * imgW + j] = image.at<cv::Vec3b>(i, j)[3 - k] / 255.0; } } } Ort::Env env(ORT_LOGGING_LEVEL_ERROR, "yolo11"); OrtCUDAProviderOptions cudaOptions; cudaOptions.device_id = 0; Ort::SessionOptions sessionOptions; sessionOptions.SetIntraOpNumThreads(1); sessionOptions.AppendExecutionProvider_CUDA(cudaOptions); sessionOptions.SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_ENABLE_ALL); static Ort::Session yolo_session(env, L"yolo11l.onnx", sessionOptions); vector<int64_t> inputShape{ 1, 3, imgH, imgW }; Ort::MemoryInfo memoryInfo = Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeCPU); Ort::Value inputTensor = Ort::Value::CreateTensor<float>(memoryInfo, imgData.data(), imgData.size(), inputShape.data(), inputShape.size()); const char* inputNames[] = { "images" }; const char* outputNames[] = { "output0" }; Ort::RunOptions runOptions; vector<Ort::Value> out = yolo_session.Run(runOptions, inputNames, &inputTensor, 1, outputNames, 1); std::vector<int64_t> shape = out[0].GetTensorTypeAndShapeInfo().GetShape(); float* data = (float*)out[0].GetTensorMutableData<void>(); //data shape {1,84,8400} /////////////////////////////postprocess//////////////////////////////////////////// std::vector<CDetectBox> allBoxes, nmsBoxes; allBoxes.reserve(200); for (int w = 0; w < 8400; w++) { float maxScore = 0; int id = -1; for (int h = 4; h < 84; h++) { if (maxScore < data[h * 8400 + w]) { maxScore = data[h * 8400 + w]; id = h - 4; } } if (maxScore > 0.3) { CDetectBox box; box.id = id; box.score = maxScore; box.x1 = data[0 * 8400 + w] - data[2 * 8400 + w] / 2; box.y1 = data[1 * 8400 + w] - data[3 * 8400 + w] / 2; box.x2 = data[0 * 8400 + w] + data[2 * 8400 + w] / 2; box.y2 = data[1 * 8400 + w] + data[3 * 8400 + w] / 2; allBoxes.emplace_back(box); } } std::sort(allBoxes.begin(), allBoxes.end(), [](CDetectBox& a, CDetectBox& b) { return a.score > b.score; }); while (!allBoxes.empty()) //nms { nmsBoxes.emplace_back(allBoxes.front()); auto it = allBoxes.begin(); allBoxes.erase(it); for (auto it = allBoxes.begin(); it != allBoxes.end();) { float iou = compute_iou(nmsBoxes.back(), *it); if (iou > 0.5) { it = allBoxes.erase(it); } else { it++; } } } ///////////////////////////////////////////////////////////////////////// for (int i = 0; i < nmsBoxes.size(); i++) { int x = nmsBoxes[i].x1; int y = nmsBoxes[i].y1; int w = nmsBoxes[i].x2 - nmsBoxes[i].x1; int h = nmsBoxes[i].y2 - nmsBoxes[i].y1; printf("%d %f %f %f %f %f\n", nmsBoxes[i].id, nmsBoxes[i].score, nmsBoxes[i].x1, nmsBoxes[i].y1, nmsBoxes[i].x2, nmsBoxes[i].y2); cv::Rect rect(x, y, w, h); cv::rectangle(image, rect, cv::Scalar(0, 255, 0), 1, 8, 0); } cv::imshow("img", image); cv::waitKey(0); return 0; }
结果如下：
