yolo-rs

yolo-rs: YOLO object detection in Rust

yolo-rs runs Ultralytics YOLO object detection models in Rust with ONNX Runtime (through ort). Give it an image and an ONNX export of YOLOv8, YOLO11, YOLO12 or YOLO26, and it returns each detected object’s bounding box, label and confidence score.

See docs.rs for the latest documentation.

Supported models

CI exports each model below from Ultralytics and checks that yolo-rs finds the same people and baseball glove in the test picture with every one of them (tests/e2e.rs). It runs on every change and weekly against the newest Ultralytics release.

Model	Tested weights	Export
YOLOv8	yolov8x	yolo export model=yolov8x.pt format=onnx
YOLO11	yolo11x	yolo export model=yolo11x.pt format=onnx
YOLO12	yolo12x	yolo export model=yolo12x.pt format=onnx
YOLO26	yolo26x	yolo export model=yolo26x.pt format=onnx end2end=False

The smaller sizes of each version (n, s, m, l) have the same inputs and outputs. YOLO26 needs end2end=False for now: its default end-to-end export has a different output (#90).

Features

• Object detection: bounding boxes, labels and confidence scores, with non-maximum suppression.
• COCO labels built in: the 80 classes the official YOLO detection models are trained on; bring your own labels for custom models.
• ONNX Runtime: runs on the CPU or any execution provider ort supports, such as CoreML or CUDA.
• Pure Rust API: takes an image DynamicImage.

Performance

CI benchmarks every tested model on a GitHub-hosted Arm runner (ubuntu-26.04-arm, CPU only) with criterion (cargo bench --bench inference):

Model	Inference (640×640)
YOLOv8x	1.37 s
YOLO11x	1.09 s
YOLO12x	1.24 s
YOLO26x	1.08 s

Preparing the input (resizing and normalising the image) takes about 11 ms.

On a MacBook Pro (2024) with M3 Max, the median of 20 runs after 5 warm-up runs:

Model	CoreML	CPU
YOLOv8x	28 ms	374 ms
YOLO11x	45 ms	322 ms
YOLO12x	164 ms	373 ms
YOLO26x	47 ms	345 ms

Times cover inference (the model and non-maximum suppression) on the test picture, with ONNX Runtime from ort 2.0.0-rc.12 and models exported by Ultralytics 8.4.51.

Building on macOS

ONNX Runtime’s Apple build calls __isPlatformVersionAtLeast, which lives in the compiler runtime that Rust does not link by default. Without it, linking a binary, test or benchmark that uses yolo-rs fails with Undefined symbols for architecture arm64: "___isPlatformVersionAtLeast". Add this to your project’s .cargo/config.toml:

[target.'cfg(target_vendor = "apple")']
rustflags = ["-Clink-arg=-fapple-link-rtlib"]

Or emit it from your binary’s build script: println!("cargo:rustc-link-arg=-fapple-link-rtlib");.

CoreML by default

ort lets ONNX Runtime pick a device for each session, and on macOS it always adds CoreML, even when you register only the CPU execution provider. from_filename_v8 therefore runs on CoreML on a Mac. To run on the CPU, build the session yourself with AutoDevicePolicy::PreferCPU and hand it to new_v8:

use ort::session::Session;
use ort::session::builder::AutoDevicePolicy;
use yolo_rs::model::YoloModelSession;

let environment = ort::init().build()?;
let session = Session::builder(&environment)?
    .with_auto_device(AutoDevicePolicy::PreferCPU)?
    .commit_from_file("yolo11x.onnx")?;
let model = YoloModelSession::new_v8(session);

Examples

• yolo-cli: Command-line interface for running YOLO inference on images.

Acknowledgements

This project is inspired by the YOLOv8 example from the ort repository.

License

This project is licensed under the Apache License 2.0.