Different result of ultralytics and onnxruntime

[alireza00bin]

Hi, I trained a model using YOLOv11 and then export .onnx format. When I use ultralytics library to do detection, it did well and detect all the classes but when I want to use onnx format using onnxruntime, it did not detect some classes:


the first image is the result of ultralytics library and the other one is the result of onnxruntime.
For example in onnxruntime result, it did not detect the bike object while it has a high confidence in ultralytics result

[UltralyticsAssistant]

👋 Hello @alireza00bin, thank you for your interest in Ultralytics 🚀! We recommend you check out our Docs for detailed examples and common questions. You might find specific sections on Python and CLI usage very helpful.

If this is a 🐛 Bug Report, we'd appreciate it if you could provide a minimum reproducible example, which will greatly aid us in debugging this issue. It’s very helpful to see the specific inputs and outputs that are causing the problem. As an automated response, an Ultralytics engineer will also be here to assist you soon.

Additionally, for questions relating to custom training or model export issues, please try to include as much relevant information as possible, such as training logs, configuration details, and any modifications made to the model or dataset.

Our community is a great resource if you're looking to discuss ideas or need further assistance. You can join us on Discord 🎧 for live discussions, share your experiences on our Subreddit, or participate in more structured conversations on Discourse.

Upgrade

Please make sure you are using the latest ultralytics package with all dependencies updated. You can upgrade in a Python>=3.8 environment with PyTorch>=1.8:

pip install -U ultralytics

Environments

You can test your setup on various supported environments:

• Notebooks with free GPU:
• Google Cloud Deep Learning VM or Amazon Deep Learning AMI: Check our GCP Quickstart Guide and AWS Quickstart Guide
• Docker Image: See Docker Quickstart Guide

Status

Our continuous integration (CI) system checks that all the features work correctly across platforms daily. If the badge above is green, it indicates that all tests are currently passing.

We appreciate your patience and report! Your observations and inputs are invaluable for improving our tools.

[glenn-jocher]

@alireza00bin it sounds like there might be a discrepancy in how the ONNX model is being processed by ONNX Runtime. Ensure that the ONNX export was successful and that ONNX Runtime is configured correctly. You might also want to check if the model's input preprocessing and confidence thresholds are consistent between the two environments. For further guidance, you can refer to the ONNX export documentation.