ros-yolov5-node

For ROS2, uses ONNX GPU Runtime (Python SDK) to inference YOLOv5.

Comparisons

• https://github.com/ms-iot/ros_msft_onnx: seems hardcoded to do post/pre-processing for specifically Tiny-YOLOv2 and caters to only object detection and object pose. Unfortunately isn't a library for running ONNX models in general then supplying your own pre/post-processing code.
• https://github.com/raghavauppuluri13/yolov5_pytorch_ros: requires Conda so is not rosdep compatible, plus PyTorch is huge. Does not support ROS2.
• https://github.com/mats-robotics/yolov5_ros: clones YOLOv5 from source as a submodule then needs to install its requirements.txt so not rosdep compatible. Does not support ROS2 either.

Interesting

Rename to onnx_yolo_ros?

Why? Because implementation language & deep learning framework aside, the pre & post-processing code for YOLO models are largely the same between variants. Hence, once the model is exported/standardized to ONNX, it is possible to just swap which YOLO model you want to use. For example, from YOLOv5 to YOLOv4 or even YOLOv7.

That said, I have yet to thoroughly test the above. Even more so given YOLOv7 heavily uses YOLOv5's codebase. Separate implementations might be needed for different variant "generations".

Note

Whenever you see size or shape, it has been converted to be (width, height). Be mindful that numpy images are still (height, width, channel), so you might want to apply img.shape[1::-1].