// the find
enazoe/yolo-tensorrt
TensorRT8.Support Yolov5n,s,m,l,x .darknet -> tensorrt. Yolov4 Yolov3 use raw darknet *.weights and *.cfg fils. If the wrapper is useful to you,please Star it.
A C++ wrapper around NVIDIA's DeepStream reference YOLO TensorRT code that builds engines for Yolov3, Yolov4 and Yolov5 (n through x, plus p6) from darknet .weights/.cfg files or converted PyTorch exports. It is aimed at people putting detection on Jetson boards or x86 Windows and Linux machines who want a small Detector class instead of wiring TensorRT by hand.
- FP32, FP16 and INT8 are all selected through one Config struct, and INT8 calibration is driven by a plain text list of images, so switching precision is a config change rather than a separate pipeline.
- Batch inference and non-square input sizes are both implemented, which many forks of this code skip.
- The public surface is small: a pimpl Detector with init() and detect() over std::vector<cv::Mat>, so TensorRT types stay out of the caller's headers.
- The benchmark tables split detect time from inference time and list Jetson NX numbers per precision, which makes the figures easier to check against your own hardware.
- The build instructions still target TensorRT 7.1.3.4, VS2015 and CUDA 11.0, and the sln/props directory is full of version-pinned .props files. Anyone on a current TensorRT has to work out the migration on their own.
- The Yolov5 path is manual. You convert the yaml to a darknet-style cfg with scripts/yaml2cfg.py and bring your own weights, so any architecture change upstream in Ultralytics means repeating that step.
- Dynamic input size is still an unchecked item on the feature list, so each engine is built for one shape. The Config default detect_thresh of 0.9 is far above the usual 0.25 to 0.5 range, so the detector will look broken to anyone who runs it without reading the config.
- There is no test directory or CI configuration in the tree, and the benchmark numbers come from JetPack 4.4.1 and a 1080 Ti. Recent push activity does not show that the newer TensorRT and JetPack versions have been exercised.