Understanding Yolov8 Task Input and Output Formats in OnnxRuntime

[FunJoo]

Hello, author. Recently, while using your library for image processing, I've been exploring methods to improve performance. By reviewing code from others on GitHub, I found that using floating-point images and the Threshold method instead of iterating through pixels can indeed enhance speed. For example, OpenCVSharp supports this functionality. The Halcon library, which I use at work, also supports this feature.

So, I convert the image to Tensor<float> using Halcon, and after obtaining the result from OnnxRuntime, I use Halcon again to transform the Tensor<float> result back into a floating-point image for further processing. Following your source code, I successfully replaced ImageSharp with Halcon in the Detection task and achieved a significant performance improvement. However, I encountered several challenges when converting the Segmentation code.

Therefore, I would like to seek your advice on where I can learn about the Yolov8 task's input and output formats in OnnxRuntime. I aim to complete the replacement for the remaining tasks such as segmentation, pose estimation, and classification.

Once again, thank you for your open-source code.

[dme-compunet]

The place to understand the output formats of YOLOv8 is in the official repo: https://github.com/ultralytics/ultralytics it is scattered with some discussions and issues, focus there. the code of the current repo can also help you figure it out.