rtdetr_rust

2024 SOTA object detector RT-DETR (CVPR2024, https://github.com/lyuwenyu/RT-DETR) inference using Rust + Ort

Currently, I only test ONNX-runtime with CPU backend, without optimization. The latency (including image resize and neural network inference) is about 565ms.

Get start

• Setup Rust compilier and compile the code

  cargo build --release
  

• Prepare RT-DETR ONNX and an image to detect objects (please follow https://github.com/lyuwenyu/RT-DETR/tree/main/rtdetrv2_pytorch#usage)
• Download libonnxruntime (please choose the right library according to your OS and platform. Here I use Ubuntu on x86_64 without GPU as an exaxmple):

  wget https://github.com/microsoft/onnxruntime/releases/download/v1.19.0/onnxruntime-linux-x64-1.19.0.tgz
  tar xzvf onnxruntime-linux-x64-1.19.0.tgz
  ln -s $(pwd)/onnxruntime-linux-x64-1.19.0/lib/libonnxruntime.so.1.19.0 libonnxruntime.so
  

• Run

  ORT_DYLIB_PATH=../../libonnxruntime.so ./target/release/rtdetr_rust path/to/your/image
  

Reference

• https://github.com/lyuwenyu/RT-DETR
• https://github.com/pykeio/ort