yolo-tensorrt

INTRODUCTION

the project is the encapsulation of tensorrt yolo implementation. The origin code is here. And you must have the trained yolo model(.weights) and .cfg file from the darknet.

PLATFORM

model	gpu	precision	memory used	inference time
yolov3-416x416	gtx1050	INT8		25ms
yolov3-416x416	gtx1050	FLOAT32		50ms
yolov3-608x608	gtx1050	INT8	~450M	~50ms
yolov3-608x608	gtx1050	FLOAT32	~1000M	~95ms
yolov3-416x416	jetson nano (15w)	HALF(FP16)		250ms

WRAPPER

user-friendly

Detector detector;
Config config;
config.inference_precison = INT8;
detector.init(config);

cv::Mat mat_image = cv::imread("dog.jpg");
std::vector<Result> res;
detector.detect(mat_image, res)

How to use yolo-trt as DLL and SO libraries

download the pretrained model

git clone https://github.com/enazoe/yolo-tensorrt.git
cd scripts/
python3 download.py
//python3.exe download

windows10

requirements : cuda 10.0 , cudnn 7.5 , TensorRT 5.1.5 , opencv3.3 , gflags , vs2015
build:

open sln/sln.sln file
- dll project : the interface of the trt yolo lib
- demo project : test of the dll

ubuntu

jetson nano

requirements : gflags , JetPack 4.2.2

make

sudo apt-get install libgflags-dev
cd yolo-tensorrt/
mkdir build
cd build/
cmake ..
make

the code generate the libdetector.so lib, and the test code

note: when the platform is jetson nano the gencode arch must be set compute_53,code=sm_53

API

struct Config
{
	std::string file_model_cfg = "configs/yolov3.cfg";

	std::string file_model_weights = "configs/yolov3.weights";

	float detect_thresh = 0.9;

	ModelType net_type = YOLOV3;

	Precision inference_precison = INT8;
	
	int gpu_id = 0;

	std::string calibration_image_list_file_txt = "configs/calibration_images.txt";
};

class API Detector
{
public:
	explicit Detector();
	~Detector();

	void init(const Config &config);

	void detect(const cv::Mat &mat_image, std::vector<Result> &vec_result);

private:
	Detector(const Detector &);
	const Detector &operator =(const Detector &);
	class Impl;
	Impl *_impl;
};

intjun/yolo-tensorrt