Finally, a YOLO runtime that doesn't need Python

What it does YoloDotNet is a C# inference library for modern YOLO families — from YOLOv5u up to YOLOv26, plus YOLO-World and RT-DETR. It handles classification, detection, oriented bounding boxes, segmentation, and pose estimation on images or video streams. The library sits on top of ONNX Runtime and SkiaSharp, deliberately leaving OpenCV and any Python interpreter out of the dependency tree.

The interesting bit Instead of one bloated package, execution providers are split into separate NuGet packages — CPU, CUDA/TensorRT, OpenVINO, CoreML, DirectML — so you only ship the native dependencies you actually need. The author treats preprocessing as configuration rather than magic, forcing you to match resize modes and thresholds to how the model was trained. That pedantry is the point: it keeps accuracy predictable in production.

Key highlights

• Supports five vision tasks — classification, object detection, OBB, segmentation, and pose estimation — across a wide range of YOLO model versions
• Modular execution providers avoid native runtime conflicts; you reference exactly one hardware backend
• Runs on Windows, Linux, and macOS depending on provider (CPU is universal; CUDA and OpenVINO skip macOS; CoreML is Apple-only; DirectML is Windows-only)
• Version 4.2 adds ROI inference, JSON and YOLO-format export helpers, and contour extraction for segmented objects
• Users must supply their own ONNX models; the project does not bundle pretrained weights

Caveats

• Most Ultralytics YOLO models carry AGPL-3.0 or commercial licensing; the MIT-licensed library does not transfer any rights to the weights you bring
• Preprocessing settings must match the training configuration; the README warns that accuracy is not automatic and points to a separate accuracy guide
• Only one execution provider can be referenced at a time; mixing them causes native runtime conflicts

Verdict Worth a look if you are building desktop apps, backend services, or real-time pipelines in .NET and want deterministic vision inference with explicit control over preprocessing and memory. Skip it if you need training tools or a quick Python script — this is strictly an inference engine.