最近實現了YOLOv10的tensorRT推理代碼除了后處理部分只適合YOLOv10之外，其余部分基本可以在yolo系列通用~學習記錄~。

#include <fstream>
#include <iostream>
#include <vector>
#include <opencv2/opencv.hpp>
#include "NvInfer.h"
#include "cuda_runtime_api.h"
#include "preprocess.hpp"
#include "postprocess.hpp"
#include "packageimage.hpp"
using namespace cv;
using namespace std;
using namespace nvinfer1;const int INPUT_H = 640;
const int INPUT_W = 640;
const int NUM_CLASSES = 5;
const float CONF_THRESH = 0.5;
const float IOU_THRESH = 0.5;
const int Inputsize_w = 640;
const int Inputsize_h = 200;
const int batch_size = 7;// 替換為你的類別名
std::vector<std::string> class_names = { };// Logger 類（TensorRT 要求）
class Logger : public ILogger {
public:void log(Severity severity, const char* msg) noexcept override {if (severity <= Severity::kINFO)std::cout << "[TRT] " << msg << std::endl;}
} gLogger;// 安全 CUDA 分配
void* safeCudaMalloc(size_t size) {void* ptr = nullptr;cudaMalloc(&ptr, size);return ptr;
}int main() {std::string engine_path = "";std::string image_path = "";std::string imagesave_path = "";// 讀取 engine 文件std::ifstream file(engine_path, std::ios::binary);if (!file) {std::cerr << "Failed to open engine file.\n";return -1;}file.seekg(0, file.end);size_t engine_size = file.tellg();file.seekg(0, file.beg);std::vector<char> engine_data(engine_size);file.read(engine_data.data(), engine_size);file.close();// 創建 runtime 和 engineIRuntime* runtime = createInferRuntime(gLogger);ICudaEngine* engine = runtime->deserializeCudaEngine(engine_data.data(), engine_size);IExecutionContext* context = engine->createExecutionContext();// 獲取輸入輸出索引int inputIndex = engine->getBindingIndex("images");//int outputIndex = engine->getBindingIndex(engine->getBindingName(1)); // 假第二個綁定是輸出int outputIndex = engine->getBindingIndex("output0");std::vector<Patches> batches;package(batches, image_path, Inputsize_w, Inputsize_h, INPUT_H, INPUT_W);// 圖像預處理/*cv::Mat img = cv::imread(image_path);cv::Mat blob;float ratio;int pad_w, pad_h;preprocess(img, blob, ratio, pad_w, pad_h, INPUT_W, INPUT_H);*/// 分配 Host 和 Device 內存size_t input_size = batch_size * 3 * INPUT_H * INPUT_W * sizeof(float);size_t output_size = batch_size * 300 * 6 * sizeof(float); // 假設最多 300 個目標，每個6個值float* input_host = new float[batch_size * 3 * INPUT_H * INPUT_W];float* output_host = new float[batch_size * 300 * 6];for (int i = 0; i < batch_size; i++){memcpy(input_host + i * 3 * INPUT_H * INPUT_W, batches[i].patch.ptr<float>(), 3 * INPUT_H * INPUT_W * sizeof(float));}float* input_device = (float*)safeCudaMalloc(input_size);float* output_device = (float*)safeCudaMalloc(output_size);void* bindings[2];bindings[inputIndex] = input_device;bindings[outputIndex] = output_device;cudaMemcpy(input_device, input_host, input_size, cudaMemcpyHostToDevice);context->setBindingDimensions(inputIndex, Dims4(batch_size, 3, INPUT_H, INPUT_W));context->enqueueV2(bindings, 0, nullptr);cudaMemcpy(output_host, output_device, output_size, cudaMemcpyDeviceToHost);// 原圖用于繪制結果cv::Mat img = cv::imread(image_path);std::cout << "batch_size = " << batch_size << ", batches.size() = " << batches.size() << std::endl;for (int i = 0; i < batch_size; i++) {std::vector<Detection> results;float* cur_output = output_host + i * 300 * 6;postprocess(cur_output, 300, CONF_THRESH, results,batches[i].pad_w, batches[i].pad_h, batches[i].ratio);std::cout << "Patch " << i << " detected " << results.size() << " objects." << std::endl;for (const auto& det : results) {int x = det.bbox.x + batches[i].x_offset;int y = det.bbox.y + batches[i].y_offset;int w = det.bbox.width;int h = det.bbox.height;cv::rectangle(img, cv::Point(x, y), cv::Point(x + w, y + h), cv::Scalar(0, 255, 0), 2);std::string label = class_names[det.class_id] + " " + std::to_string(det.conf);cv::putText(img, label, cv::Point(x, y - 5), 0, 0.5, cv::Scalar(255, 255, 255), 1);}}cv::imwrite(imagesave_path, img);std::cout << "Result saved as: " << imagesave_path << std::endl;// 清理資源delete[] input_host;delete[] output_host;cudaFree(input_device);cudaFree(output_device);context->destroy();engine->destroy();runtime->destroy();return 0;
}

預處理：

//頭文件:
// preprocess.hpp
#pragma once
#include <opencv2/opencv.hpp>/*** @brief 將原始圖像預處理為 TensorRT 模型輸入格式** @param img      輸入圖像（原圖）* @param output   輸出張量，格式為 [3, INPUT_H, INPUT_W]，float32，CHW* @param scale    返回：縮放比例* @param pad_w    返回：左側 padding 寬度* @param pad_h    返回：上方 padding 高度* @param new_w    模型輸入寬度（例如 640）* @param new_h    模型輸入高度（例如 640）*/
void preprocess(const cv::Mat& img, cv::Mat& output, float& scale, int& pad_w, int& pad_h, int new_w, int new_h);//cpp:
#include "preprocess.hpp"
#include <opencv2/opencv.hpp>void preprocess(const cv::Mat& img, cv::Mat& output, float& scale, int& pad_w, int& pad_h, int new_h, int new_w){ int orig_w = img.cols;int orig_h = img.rows;scale = std::min((float)new_w / orig_w, (float) new_h / orig_h);int resized_w = int(scale * orig_w);int resized_h = int(scale * orig_h);// 縮放圖像cv::Mat resized;cv::resize(img, resized, cv::Size(resized_w, resized_h));cv::Mat padded(new_h, new_w, CV_8UC3, cv::Scalar(114, 114, 114));pad_w = (new_w - resized_w) / 2;pad_h = (new_h - resized_h) / 2;resized.copyTo(padded(cv::Rect(pad_w, pad_h, resized_w, resized_h)));cv::Mat float_img;padded.convertTo(float_img, CV_32FC3, 1.0 / 255.0);// HWC -> CHWstd::vector<cv::Mat> channels(3);cv::split(float_img, channels);output = cv::Mat(3, new_h * new_w, CV_32F);// 將每個通道拉平并寫入 outputfor (int i = 0; i < 3; ++i) {memcpy(output.ptr<float>(i), channels[i].data, new_h * new_w * sizeof(float));}}

后處理

//頭文件：
#pragma once
#include <vector>
#include <opencv2/opencv.hpp>struct Detection {cv::Rect bbox;    // 檢測框int class_id;     // 類別編號float conf;       // 置信度
};/*** @brief 對模型輸出進行后處理：置信度篩選 + NMS** @param output      TensorRT 輸出（通常是 [N, 6]）* @param num_preds   預測框數量（例如：1000）* @param conf_thresh 置信度閾值* @param out_dets    輸出：保存最終保留的檢測結果*/
void postprocess(const float* output, int num_preds, float conf_thresh, std::vector<Detection>& out_dets,int pad_w, int pad_h, float ratio);//cpp:
#include "postprocess.hpp"
#include <opencv2/opencv.hpp>
void postprocess(const float* output, int num_preds, float conf_thresh, std::vector<Detection>& out_dets, int pad_w, int pad_h, float ratio) {std::vector<Detection> raw_dets;for (int i = 0; i < num_preds; i++) {const float* det = output + i * 6;float x1 = det[0], y1 = det[1], x2 = det[2], y2 = det[3];float conf = det[4];int class_id = static_cast<int>(det[5]);if (conf < conf_thresh) {continue;}x1 = (x1 - pad_w) / ratio;y1 = (y1 - pad_h) / ratio;x2 = (x2 - pad_w) / ratio;y2 = (y2 - pad_h) / ratio;int left = static_cast<int>(x1);int top = static_cast<int>(y1);int right = static_cast<int>(x2);int bottom = static_cast<int>(y2);int width = right - left;int height = bottom - top;cv::Rect bbox(static_cast<int>(left),static_cast<int>(top),static_cast<int>(width),static_cast<int>(height));std::cout << "x1 = " << x1 << ", y1 = " << y1 << ", x2 = " << x2 << ", y2 = " << y2 << std::endl;out_dets.push_back(Detection{ bbox, class_id, conf });}
}