TensorRT-YOLO

Why we included this project

Deploying a YOLO model on NVIDIA hardware means more than exporting weights. You still have to convert to TensorRT engines, speed up preprocessing, and hook the results into an application. This toolkit bundles that whole pipeline into C++ and Python APIs, with CUDA kernels and CUDA Graph execution doing the speed work and TensorRT plugins handling the postprocessing. One code path serves detection, instance segmentation, pose estimation, classification, and rotated bounding boxes, so a team can standardize on a single deployment library instead of writing per-task glue code. Embedded setups get attention too, with x86 and ARM builds, zero-copy memory handling for Jetson-style devices, and a single-header C++ option that keeps CUDA and TensorRT link dependencies out of larger applications. Runnable examples for each task make it a handy reference when you are wiring up your own inference service or comparing engine settings.