Using personal image classifier for traffic sign recognition

[Svein-Tore]

HI!
I'm trying to use the image classifier to get a car turn left or right. The sign I want to recognize are:

I have tried to use both new and old image recognizer. But it says right nearly all the time.

Any suggestions?

[SteveJG]

@Svein-Tore

But it says right nearly all the time.

Which is the highest probability? How close are the highest two selections? Are the images you provide for analysis more or less the same but with background clutter (you examples have no background). See Personal image classifier display .

• Perhaps you need to enter more training images
• or the PIC algorithm can not distinguish a small, relatively simple differences in the sign. What do you expect for free? Perfect recognition is probably impossible.
• you might have better luck with TMIC . Try TMIC extension and let us know.

TMIC: App Inventor Extension for the Deployment of Image Classification Models Exported from Teachable Machine

TMIC: App Inventor Extension for the Deployment of Image Classification Models Exported from Teachable Machine Version: 1.0 Released: August, 25 2022 Tested Android 9, 12, 13, companion and compiled TMIC is an App Inventor extension for the deployment of ML models for image classification developed with Google Teachable Machine in educational settings. Google Teachable Machine is an intuitive visual tool that provides workflow-oriented support for the development of ML models for image clas…

[Svein-Tore]

Thanks @SteveJG, I will test this extension

[Svein-Tore]

So I have tested this extension, and it is clearly better. The best is the possibility to take many pictures by holding down the key. MITs program can only shoot one picture at a time. Taking i.e. some hundred pictures is done very fast. And the model is even better. I had to change the backgroundcolor of one of the picture from blue to yellow though. The car is also starting to look like something:

This is as you can see 3D printed (yes I have to do something with the wires🤣). I am using a Micro:bit and some bluetooth extensions. One thing I miss in TMIC is that the result could have been sorted with regard to probability (@cgvw ) (the one with highest probability as the first in the dictionary). I have done that with a procedure. I think this will be a cool project for my students.