Do I Need To Train My Own Analog?

[whitedavidp]

First, I have an very old meter (circa mid-1960's). It was likely inexpensive and low-grade to begin with. And time has not made it better. Our water district has no plans to update it any time soon and will not allow me to purchase a meter of my choice. So I am stuck with it.

I have done lots of work trying to obtain better images. But it seems the issue is the plastic or glass cover over the dial coupled with, perhaps, some moisture on the inside of the cover. The outside is clean. As you can see, the image looks nothing like those seen in the included models. But using both zoom and negative image capture modes, I am getting at least decent digit recognition.

But the analog readings are way off and I can't seem to impact them doing anything. I would have thought that these would be easiest to make out and read with some accuracy. I have no idea how the analog models work. But with a dial, it seems that if you make out the pointer tip/end and its angle relative to something - in my case it is just the vertical axis, you should be able to infer the reading on a 0-9 scale. But that is just me wondering.

So do I need to try and train my own model for this? Any suggestions are welcome. Thanks

[SybexX]

Try these models SybexXx.zip

[SybexX]

How did you set your ROIs? https://jomjol.github.io/AI-on-the-edge-device-docs/ROI-Configuration/

[whitedavidp]

Very carefully. I even removed the synchronized box and tried to be extra precise. It may be settling-in as I twiddle the LED a bit more:

[SybexX]

So it should look roughly

The height of your inner rectangle should be as high as your 1 and the width as wide as your 0 (height = the longest number, width = the widest number).

[whitedavidp]

Thanks for all of your help. I have learned a bunch of stuff here. Sorry for taking so much time. I will try to do the same for others.

This is all, quite frankly, amazing! The quality and thought that went into this, even just having the original idea that it all could be done on the camera itself, and that it actually can work on that tiny platform is just so cool (and unexpected). Best!