YOLOV11 and export to blob format performance

[wokka77]

Newbie question -

I have a model I created in YOLOv11 and its helped with performance so far.

I want to export it as .blob format to run it on an OAK camera.

If I export the model as blob format, will the conversion process alter the performance of the YOLOv11 model at all?

My apologies, am not across the finer points of these models....

Thanks in advance.

[UltralyticsAssistant]

👋 Hello @wokka77, thank you for your interest in Ultralytics 🚀! We recommend checking out the Docs where you'll find extensive information, including Python usage and exporting models that may help answer your questions.

If this is a 🐛 Bug Report, please provide a minimum reproducible example to assist us in debugging it efficiently.

If this is a general ❓ Question or relates to custom training and exporting, please share as much context as possible regarding your workflow. This includes any export logs, configurations, and steps you've already completed, as well as details about the OAK camera system you are using.

Additionally, here are a few ways to connect with our community for more support:

• Real-time chat: Discord 🎧
• In-depth discussions: Discourse
• Community sharing: Subreddit

Upgrade

To ensure compatibility and optimal performance, consider upgrading to the latest version of the ultralytics package, which includes all requirements. Ensure you're using a Python>=3.8 environment with PyTorch>=1.8:

pip install -U ultralytics

Exporting Models to .blob Format

For exporting models, YOLO supports conversion to diverse formats. Refer to the Export Guide for detailed instructions on supported formats, including .blob format. If you've encountered specific issues during your export process, please share those details to help us better assist you.

You can also verify our supported environments:

• Google Colab:
• Paperspace Gradient:
• Docker: See Docker Quickstart

Status

If this badge is green, all Ultralytics CI tests are currently passing. CI tests validate correct functionality of YOLO modes and tasks across platforms daily.

🤖 This is an automated response, but no worries! An Ultralytics engineer will assist you soon. 😊

[glenn-jocher]

@wokka77 for .blob format conversion with OAK devices, the export process optimizes the model specifically for Intel Myriad X VPU hardware. This conversion typically maintains detection accuracy while significantly improving inference speed on OAK cameras, but you should validate performance using our Benchmark mode (https://docs.ultralytics.com/modes/benchmark) before deployment. The Ultralytics export docs (https://docs.ultralytics.com/modes/export) include guidance for hardware-specific optimizations.

[wokka77]

Hi,

Thank you for the response.

Can I create the .blob file out of Ultralytics and immediately copy blob file onto the OAK camera without needing to do anything else? And would you suggest we enable int8 quantization for the conversion or any other options to squeeze the max out of performance please?

Luxonis appear to also use OpenVINO 2021.4 format in their conversions.

I'm just trying to work out whether we would see a difference between the two conversion processes.

I do like the export straight out of Ultralytics, thats a really useful function to have.

I will run some benchmark tests as well.

Thanks.

[glenn-jocher]

@wokka77 for OAK deployments using .blob format, we recommend first exporting to OpenVINO format with format='openvino' then using Luxonis's blobconverter tool for final optimization. Enable INT8 quantization if your OAK device supports it, but validate accuracy changes using our Benchmark mode. The Ultralytics OpenVINO export matches modern implementations, but you should cross-check performance against Luxonis's 2021.4-based workflow.