01 / OVERVIEW
Overview
SimdYolo implements YOLO inference and supported detection-head fine-tuning in pure C#. It reads supported PT and ONNX files, uses SIMD / intrinsics on the CPU, and offers Vulkan and Metal backends. The repository includes a web example, model documentation, and reproducible validation reports.
02 / CAPABILITIES
Key capabilities
01
Load supported PT and ONNX model files directly
02
Object detection, instance segmentation, and pose estimation
03
CPU SIMD, Vulkan / Metal backends, and detection-head fine-tuning
03 / POSSIBILITIES
Ideas to explore
- Explore detection, segmentation, and pose in .NET image applications
- Study model loading and execution without a Python inference runtime
- Experiment with labeling, pre-labeling, and supported detection-head fine-tuning
04 / GET STARTED
Getting started
Before you begin
- A .NET 10 SDK or a matching runtime
- Supported weights and tasks, with model licenses reviewed
- Compatible GPU drivers when using a GPU backend; validation differs by backend
Get the project
git clone https://github.com/donet-org-cn/SimdYolo.gitNext steps
- Install the inference package or build from source.
- Prepare weights from the supported model list and read the inference examples.
- Validate a single image before exploring the web demo or fine-tuning workflow.
05 / SCOPE & STATUS
Current boundaries
Model and operator support is limited to the repository's compatibility list. Metal has not been validated on real hardware. Training fine-tunes the detection head with a frozen backbone.
This introduction is based on the public README. Consult the repository for current versions, licenses, and supported features.
Updated 2026-10-27