wip: C++ 参考实现（待编译验证）

deploy/cpp/main.cpp

@@ -0,0 +1,49 @@
+// 演示与对拍工具：读一张原始像素转储，推理并打印框。
+//
+//   ./film_predict <model.xml> <image.bin> <宽> <高> <通道> [conf] [tile]
+//
+// 输入是连续的 HWC uint8 缓冲（用 scripts/dump_raw.py 从图片生成）。产线里通常直接
+// 拿相机帧的指针，不需要这一步。tile 传 0 表示图本身就是网络输入尺寸。
+#include "film_detector.hpp"
+
+#include <cstdio>
+#include <fstream>
+#include <iostream>
+#include <vector>
+
+int main(int argc, char** argv) {
+    if (argc < 6) {
+        std::cerr << "用法: " << argv[0]
+                  << " <model.xml> <image.bin> <宽> <高> <通道> [conf] [tile-overlap]\n";
+        return 1;
+    }
+    const std::string model_path = argv[1], image_path = argv[2];
+    const int width = std::stoi(argv[3]), height = std::stoi(argv[4]);
+    const int channels = std::stoi(argv[5]);
+    const float conf = argc > 6 ? std::stof(argv[6]) : 0.25f;
+    const int overlap = argc > 7 ? std::stoi(argv[7]) : 32;
+
+    std::ifstream file(image_path, std::ios::binary);
+    if (!file) { std::cerr << "打不开 " << image_path << "\n"; return 1; }
+    std::vector<unsigned char> pixels((std::istreambuf_iterator<char>(file)),
+                                       std::istreambuf_iterator<char>());
+    const size_t expected = static_cast<size_t>(width) * height * channels;
+    if (pixels.size() != expected) {
+        std::cerr << "缓冲大小 " << pixels.size() << " 与 " << width << "x" << height
+                  << "x" << channels << " 不符（应为 " << expected << "）\n";
+        return 1;
+    }
+
+    film::Detector detector(model_path);
+    film::ImageView image{pixels.data(), width, height, channels};
+    const auto boxes = (width == detector.input_size() && height == detector.input_size())
+        ? detector.infer_tile(image, conf)
+        : detector.infer(image, conf, 0.5f, overlap);
+
+    std::printf("%zu\n", boxes.size());
+    for (const auto& b : boxes) {
+        std::printf("%d %.6f %.6f %.6f %.6f %.4f\n",
+                    b.class_id, b.x1, b.y1, b.x2, b.y2, b.score);
+    }
+    return 0;
+}