ssd mobilenet v2 coco 2018

.gitignore

@@ -1,3 +1 @@
 ssd_mobilenet_v2_coco_2018_03_29/demo
-
-**/demo

ssd_mobilenet_v2_coco_2018_03_29/README.md

@@ -19,18 +19,22 @@ python demo.py --model ssd_mobilenet_v2_coco_2018_03_29_2026jul.onnx --image exa
 ```
 
 ### C++
-The C++ demo runs inference with OpenCV's DNN module. Adjust the OpenCV paths to your setup:
+The C++ demo runs inference with ONNX Runtime (C++ API) and uses OpenCV only for image I/O.
+Install ONNX Runtime (C++) from https://github.com/microsoft/onnxruntime/releases — this build
+uses `onnxruntime-linux-x64-1.25.0` — and adjust the ONNX Runtime and OpenCV paths to your setup:
 ```bash
+ORT=/path/to/onnxruntime-linux-x64-1.25.0    # ONNX Runtime release dir (contains include/ and lib/)
 OCV=/path/to/opencv                          # OpenCV source tree
 OCVBUILD=/path/to/opencv/build               # OpenCV build directory (generated headers + libs)
 g++ -std=c++17 demo.cpp -o demo \
+  -I$ORT/include \
   -I$OCV/include \
   -I$OCV/modules/core/include \
-  -I$OCV/modules/dnn/include \
   -I$OCV/modules/imgproc/include \
   -I$OCV/modules/imgcodecs/include \
   -I$OCVBUILD \
-  -L$OCVBUILD/lib -Wl,-rpath,$OCVBUILD/lib -lopencv_dnn -lopencv_imgcodecs -lopencv_imgproc -lopencv_core
+  -L$ORT/lib -Wl,-rpath,$ORT/lib -lonnxruntime \
+  -L$OCVBUILD/lib -Wl,-rpath,$OCVBUILD/lib -lopencv_imgcodecs -lopencv_imgproc -lopencv_core
 ./demo --model ssd_mobilenet_v2_coco_2018_03_29_2026jul.onnx --image example_outputs/input_image.png --output example_outputs/output_image.png
 ```
 

ssd_mobilenet_v2_coco_2018_03_29/demo.cpp

@@ -1,4 +1,4 @@
-#include <opencv2/dnn.hpp>
+#include <onnxruntime_cxx_api.h>
 #include <opencv2/imgproc.hpp>
 #include <opencv2/imgcodecs.hpp>
 #include <array>
@@ -35,22 +35,39 @@ int main(int argc, char** argv)
     resize(rgb, rgb, Size(300, 300));
     if (!rgb.isContinuous()) rgb = rgb.clone();
 
-    int blobShape[] = {1, 300, 300, 3};
-    Mat blob(4, blobShape, CV_8U, rgb.data);
-    dnn::Net net = dnn::readNetFromONNX(model, dnn::ENGINE_ORT);
-    net.setInput(blob);
-    std::vector<String> out_str = {"detection_boxes:0", "detection_scores:0", "detection_classes:0", "num_detections:0"};
-    std::vector<Mat> outs;
-    net.forward(outs, out_str);
+    Ort::Env env(ORT_LOGGING_LEVEL_WARNING, "demo");
+    Ort::SessionOptions so;
+    Ort::Session session(env, model.c_str(), so);
+    Ort::AllocatorWithDefaultOptions alloc;
+
+    auto in_name = session.GetInputNameAllocated(0, alloc);
+    const char* in_names[] = {in_name.get()};
+
+    size_t out_count = session.GetOutputCount();
+    std::vector<Ort::AllocatedStringPtr> out_holders;
+    std::vector<std::string> out_str;
+    std::vector<const char*> out_names;
+    for (size_t i = 0; i < out_count; ++i)
+    {
+        out_holders.push_back(session.GetOutputNameAllocated(i, alloc));
+        out_str.push_back(out_holders.back().get());
+        out_names.push_back(out_str.back().c_str());
+    }
+
+    std::array<int64_t, 4> shape = {1, 300, 300, 3};
+    auto mem = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);
+    Ort::Value input = Ort::Value::CreateTensor<uint8_t>(mem, rgb.data, 300 * 300 * 3, shape.data(), shape.size());
+
+    auto outs = session.Run(Ort::RunOptions{nullptr}, in_names, &input, 1, out_names.data(), out_names.size());
 
     const float *boxes = 0, *scores = 0, *classes = 0, *num = 0;
-    for (size_t i = 0; i < out_str.size(); ++i)
+    for (size_t i = 0; i < out_count; ++i)
     {
         const std::string& n = out_str[i];
-        if (n.find("detection_boxes") != std::string::npos) boxes = (const float*)outs[i].data;
-        else if (n.find("detection_scores") != std::string::npos) scores = (const float*)outs[i].data;
-        else if (n.find("detection_classes") != std::string::npos) classes = (const float*)outs[i].data;
-        else if (n.find("num_detections") != std::string::npos) num = (const float*)outs[i].data;
+        if (n.find("detection_boxes") != std::string::npos) boxes = outs[i].GetTensorMutableData<float>();
+        else if (n.find("detection_scores") != std::string::npos) scores = outs[i].GetTensorMutableData<float>();
+        else if (n.find("detection_classes") != std::string::npos) classes = outs[i].GetTensorMutableData<float>();
+        else if (n.find("num_detections") != std::string::npos) num = outs[i].GetTensorMutableData<float>();
     }
     if (!boxes || !scores || !classes || !num)
     {

ssd_mobilenet_v2_coco_2018_03_29/demo.py

@@ -4,6 +4,7 @@ import os
 
 import cv2 as cv
 import numpy as np
+import onnxruntime as ort
 
 here = os.path.dirname(os.path.abspath(__file__))
 
@@ -22,10 +23,9 @@ def main():
 
     rgb = cv.resize(cv.cvtColor(img, cv.COLOR_BGR2RGB), (300, 300))
 
-    net = cv.dnn.readNetFromONNX(args.model, cv.dnn.ENGINE_ORT)
-    onames = ["detection_boxes:0", "detection_scores:0", "detection_classes:0", "num_detections:0"]
-    net.setInput(rgb[None].astype(np.uint8))
-    res = net.forward(onames)
+    sess = ort.InferenceSession(args.model, providers=["CPUExecutionProvider"])
+    res = sess.run(None, {sess.get_inputs()[0].name: rgb[None].astype(np.uint8)})
+    onames = [o.name for o in sess.get_outputs()]
     boxes = res[[i for i, n in enumerate(onames) if "detection_boxes" in n][0]].reshape(-1, 4)
     scores = res[[i for i, n in enumerate(onames) if "detection_scores" in n][0]].reshape(-1)
     classes = res[[i for i, n in enumerate(onames) if "detection_classes" in n][0]].reshape(-1)