Akida vs Yolo5 model, and setting confidence score threshold for object detection

Help
#1

Hi there
For my project, I would like to compare my Akida Fomo output with YoloV5. When select the YoloV5 model in the object detection section, there is an error when the model starts to train. Is there a reason for this?

Also, is it possible to change the value for probability so that only items with confidence >80 percent show just using the Edge Impulse web based platform?

Project ID: 215205 version 4

**Context/Use Case" Working on model for uni assignment

Thanks
Allan Bourke

#2

RE: Error

  • I can’t help with this error.

RE: Confidence

  • On the Model Testing page in the Studio click the 3 dots to make a button appear that will lead to setting “Set Confidence Thresholds” dialog box. Note that in FOMO Models this setting gets set into the final Deployed Model. So for example, if you are testing the boundary conditions, make sure to set this back to something more normal for your Model.

image
image863Ă—302 65.4 KB

1 Like
#3

Hi @Bourko1,

What error are you seeing when you try to train YoloV5?

#4

Hi @Bourko1
The issue is that YOLOv5 architecture is not compatible with Akida.
The error is not very clear, but as you selected the Object Detection (Images) - BrainChip Akida™ block in the Create Impulse, you should choose the model that is compatible with Akida.

We are working on a fix in the Studio to filter out custom learn blocks not compatible with Akida (or other accelerators), but it’s a part of bigger improvements in the workflow.

Matt

#5

Thanks heaps for your assistance here.

#6

Hi there Mateusz
I am still working on my university project. I am not having very good results with training to date. I have uploaded and labelled 2 sets of 100 images (Cracks and Spalls), hoping that I can detect both using object detection.
My project needs to be finalised in about 2 weeks. I see there is a Trial offer for the better version of Edge Impulse. I was wondering if I started the trial, at the end, am I still able to access all of my results and run my final model?
Kind regards
Allan Bourke

#7

Hi @Bourko1,
As I see, you have access to the Enterprise functions as your project is under a University organization that has the Enterprise tier enabled.
Regarding your issue with model performance, I think @louis may be a better person to assist with that.

Best regards,
Mateusz