[🚀🚀🚀All Yolo Tasks · All Yolo Versions · All Yolo Runtimes🚀🚀🚀]
A unified C++ toolkit for YOLO v5/v8/v11/v26/..., covering classification/detection/segmentation/pose/obb tasks with easy python-like APIs from ultralytics/ultralytics. Support All Yolo Tasks, All Yolo Versions, All Yolo Runtimes, it's time to make all in one.
- support all
Yolotasks includingclassification/detection/segmentation/pose/obb. - support all
Yoloversions includingyolov5(anchor-based)/yolov5u(anchor-free)/yolov8/yolov11/yolov26(nms-free)/more in the future, sub versions liken/s/m/l/xare also supported. - support all
Yoloinference backends(runtime) such asOpenCV::DNN/ONNXRuntime/TensorRT/OpenVINO/RKNN/CoreML/CANN/PaddlePaddle... - easy APIs to use and integrate, as simple as python APIs from
ultralytics/ultralyticslibrary. - toolkit works out of box, provide the model and set up the config parameters, go predict!
- C++ >= 17
- GCC >= 7.5
- OpenCV == 4.13
- CUDA/ONNXRuntime/TensorRT/OpenVINO/RKNN/... are optional
- run
git clone https://github.com/sherlockchou86/one-yolo.git - run
cd one-yolo && mkdir build && cd build - run
cmake .. && make -j8or clickdebugbutton to run samples directly if you have opened the project using VS Code
you must put test data(models&video&images) at the same directory as one-yolo first before runing the samples.
build options when run cmake command:
-DBUILD_WITH_ORT=ON # enable ONNXRuntime as inference backend
-DBUILD_WITH_OVN=ON # enable OpenVINO(Intel Platform) as inference backend
-DBUILD_WITH_TRT=ON # enable TensorRT(Nvidia/CUDA Platform) as inference backend
-DBUILD_WITH_RKN=ON # enable RKNN(RockChip Platform) as inference backend
-DBUILD_WITH_CML=ON # enable CoreML(Apple Platform) as inference backend
-DBUILD_WITH_PDL=ON # enable PaddlePaddle as inference backend
-DBUILD_WITH_CAN=ON # enable CANN(HuaWei Platform) as inference backend
-DBUILD_WITH_DEL=ON # enable Denglin's SDK(Denglin/登临 Platform) as inference backend
-DBUILD_WITH_CAB=ON # enable Cambricon'SDK(Cambricon/寒武纪 Platform) as inference backend
...
if you just run `cmake ..` without any options,
one-yolo will depend on OpenCV::DNN module as inference backend by default,
so OpenCV is required for one-yolo, CUDA is optional when building OpenCV from source code.
vehicle detection & tracking task using yolov8s:
#include "Yolo.h"
#include "track/YoloTracker.h"
using namespace yolo;
int main() {
/* 1. construct YoloConfig */
YoloConfig cfg;
cfg.desc = "vehicle detection task using yolov8s(custom model)";
cfg.version = YoloVersion::YOLO8;
cfg.task = YoloTaskType::DET;
cfg.target_rt = YoloTargetRT::OPENCV_CUDA;
cfg.model_path = "./vp_data/models/det_cls/vehicel_v8s-det_c6_20260205.onnx";
cfg.input_w = 640;
cfg.input_h = 384;
cfg.batch_size = 1;
cfg.num_classes = 6;
cfg.names = {"person", "car", "bus", "truck", "2wheel", "other"};
/* 2. create Yolo using YoloConfig */
auto model = Yolo(cfg);
model.info();
/* 3. construct YoloTrackConfig */
YoloTrackConfig t_cfg;
t_cfg.algo = YoloTrackAlgo::SORT;
t_cfg.iou_thresh = 0.6f;
/* 4. create YoloTracker using YoloTrackConfig */
auto tracker = YoloTracker(t_cfg);
tracker.info();
/* 5. open video and predict frames in a loop */
cv::VideoCapture cap("./vp_data/test_video/rgb.mp4");
while (cap.isOpened()) {
// collect frame
cv::Mat frame;
if (!cap.read(frame)) {
cap.set(cv::CAP_PROP_POS_FRAMES, 0);
continue;
}
// resize original image
if (frame.cols > 720) {
cv::resize(frame, frame, cv::Size(), 0.5, 0.5);
}
// predict with batch mode (batch size == 1)
auto results = model(std::vector<cv::Mat>{frame});
// track result
tracker(results[0]);
// show and print
results[0].info(); // print summary
results[0].to_json(true); // convert structured result to json and print
results[0].to_csv(true); // convert structured result to csv and print
if (results[0].show(
false, 1.0f, DrawParam(), // show annotated image & input image(640*384) & original image with unblock mode
true, true) == 27) { // exit loop if user has pressed ESC
break;
}
/*
* you can also get structured results like below:
* auto boxes = results[0].boxes(); // get bounding boxes in detection task
* auto cls_ids = results[0].cls_ids(); // get class ids in detection task
* auto confs = results[0].confs(); // get confidences in detection task
* auto labels = results[0].labels(); // get labels in detection task
* auto track_ids = results[0].track_ids(); // get track ids in detection task
* auto track_points = results[0].track_points(); // get track points in detection task
*/
}
}video result of vehicle detection & tracking using yolov8s:
yolo8_det4.mp4
json/csv output result of vechile detection & tracking using yolov8s:
json output:
[
{
"box": {
"height": 76,
"width": 33,
"x": 368,
"y": 378
},
"cls_id": 4,
"conf": 0.8655326962471008,
"label": "2wheel",
"track_id": 1
},
{
"box": {
"height": 21,
"width": 10,
"x": 647,
"y": 145
},
"cls_id": 4,
"conf": 0.8104556202888489,
"label": "2wheel",
"track_id": 37
},
{
"box": {
"height": 15,
"width": 9,
"x": 676,
"y": 137
},
"cls_id": 4,
"conf": 0.7772445678710938,
"label": "2wheel",
"track_id": 23
},
{
"box": {
"height": 14,
"width": 7,
"x": 710,
"y": 118
},
"cls_id": 4,
"conf": 0.523908257484436,
"label": "2wheel",
"track_id": 41
},
{
"box": {
"height": 14,
"width": 12,
"x": 793,
"y": 93
},
"cls_id": 3,
"conf": 0.5332302451133728,
"label": "truck",
"track_id": 44
},
{
"box": {
"height": 128,
"width": 113,
"x": 494,
"y": 369
},
"cls_id": 1,
"conf": 0.9514954090118408,
"label": "car",
"track_id": 5
},
{
"box": {
"height": 9,
"width": 13,
"x": 721,
"y": 117
},
"cls_id": 1,
"conf": 0.7941694259643555,
"label": "car",
"track_id": 25
},
{
"box": {
"height": 9,
"width": 14,
"x": 753,
"y": 116
},
"cls_id": 1,
"conf": 0.7911720871925354,
"label": "car",
"track_id": 13
},
{
"box": {
"height": 11,
"width": 13,
"x": 770,
"y": 107
},
"cls_id": 1,
"conf": 0.5813544988632202,
"label": "car",
"track_id": 42
}
]
csv output:
id,cls_id,conf,label,track_id
1,4,0.865533,2wheel,1
2,4,0.810456,2wheel,37
3,4,0.777245,2wheel,23
4,4,0.523908,2wheel,41
5,3,0.53323,truck,44
6,1,0.951495,car,5
7,1,0.794169,car,25
8,1,0.791172,car,13
9,1,0.581354,car,42
