TensorRT-Image-Classification

C++ Visual Studio solution for Image Classification using Caffe Model and TensorRT inference platform.

Tested on:

• TensorRT-5.1.5.0.Windows10.x86_64.cuda-10.0.cudnn7.5 (GA version)
• Windows 10
• Visual Studio 2017
• OpenCV 3.4.0 with CUDA support
• CUDA 10
• cuDNN 7.5.0

Usage

• Install TensorRT for Windows
• Make sure you have OpenCV with CUDA build support installed, and copy the world.dll file to \x64\Release\
• Open solution, in main.cpp edit the path of your caffe model and input image

cv::Mat img = cv::imread("Image.jpg");
std::string model = "CaffeModel/deploy.prototxt";
std::string trained = "CaffeModel/network.caffemodel";
std::string mean = "CaffeModel/mean.binaryproto";
std::string label = "CaffeModel/labels.txt";

• Check Project Properties and make sure all dependencies paths (TensorRT, CUDA, OpenCV) are correct
• Build the solution on Release mode and run
• The first time you run this program, it will take some time to build the CUDA engine, and the engine will be saved as ClassificationTRT.engine. The next time you run the program, it will load the created engine in a short time.

Notes

• If fopen related error occurs, Modify Configuration Properties -> C/C++ -> Preprocessor in the field PreprocessorDefinitions add ;_CRT_SECURE_NO_WARNINGS

Result

Finding CUDA Device
Parsing Caffe Model
Loading ClassificationTRT.engine
CUDA NO ERROR
Initialization Time : 3.646s
Classifying Image

TOP 1 Prediction
pothole : 91.274185%


TOP 5 Predictions
pothole : 91.274185%
shadow : 8.560489%
patch : 0.106978%
patchdamaged : 0.050732%
paintasphalt : 0.007003%

Classification Time : 10ms