Query about robustness of current Blob conversion tools

mikegardner

Hi Everyone, I'm enquiring about the robustness of the latest blob converter tools

I dabbled with all of the OAK (OAK 1, OAK D and the Pi ) hardware a few years ago and found it excellent quality and a great out of the box experience with the pre-trained blobs that were available from Luxonis.

However, at that time the blob conversion was quite flaky, and I lost interest in using the platform for deploying custom image sets.

I now have a commercial project that is extremely well suited to the OAK 1 or a similar custom single camera and I'd like to explore using the conversion tools again.

Can anyone testify to their experience of using performing a complete end-to-end capture, training, validation and testing and most importantly conversion to blob and deployment to the camera?

Thanks in advance

jakaskerl

mikegardner
Afaik, we don't yet have the full end-to-end.

There are training notebooks for common models: luxonis/depthai-ml-trainingtree/master/colab-notebooks
For conversion: https://blobconverter.luxonis.com/ and https://tools.luxonis.com/ (for yolo)

Should work no problem when no unsupported layers are used. This issue still remains because MX VPU does not support all operations. There are workarounds though and this should mostly be fixed on the RVC4.

Thanks,
Jaka

mikegardner

Thanks for the explanation Jaka, I've just ordered a OAK1 POE which I believe is RVC4 based. All of the conversions I tried previously were in 2021, so hopefully the tools have matured since then.

I'll create my model and follow the documentation to convert it to a blob and see how it goes.

Thanks

jakaskerl

mikegardner
Oak-1 POE is RVC2 based, but the tools have matured since then. If you have any problems, we are happy to help.

Thanks,
Jaka

mikegardner

OK Jaka

With respect to the tools - Is there any preference on the model type, eg MobileNet-SSD or Yolo, and if Yolo, which version performs best with the tools.

Thanks

jakaskerl

mikegardner
Yolo generally has better support, tools only support yolo.

mikegardner which version performs best with the tools.

They all perform the same since tools only do the conversion. However; in terms of on-device performance, Yolov6 has shown to have the highest FPS and lowest latency on RVC2. https://docs.luxonis.com/software/ai-inference/performance/

Thanks,
Jaka