Replies: 28 comments 7 replies
-
|
|
Beta Was this translation helpful? Give feedback.
-
|
|
Beta Was this translation helpful? Give feedback.
-
|
It seems to be working better in greyscale. |
Beta Was this translation helpful? Give feedback.
-
|
I had a very hard time with red digits on my meter. I ended up passing the led through a green filter. that helped alot. also tried gray scale but had best results using negative. best |
Beta Was this translation helpful? Give feedback.
-
|
Thanks! For me so far greyscale still gives the best results, however the analogue dial is struggling currently |
Beta Was this translation helpful? Give feedback.
-
|
Your images certainly look better than mine ever did (or do right now, for that matter). I too am not able to get decent results from my analog ROIs - which are nothing like yours. I finally just turned them off. My meter reads in cubit feet and I really don't require high precision in my application - which is mostly to identify hidden leaks. This eased my mind/efforts alot as it still quite a lot of fiddling to get the digit ROIs to work well enough for me! I am quite satisfied with those results now and am simply amazed by all of this. Best! |
Beta Was this translation helpful? Give feedback.
-
|
For me it is now looking better - I have changed the analogue model and slightly tweaked settings. It looks like when a digit is NaN the value from the previous round is used in the output value: https://jomjol.github.io/AI-on-the-edge-device-docs/Correction%20Algorithm/ Which makes sense. Let's see if it continues! Config attached for reference!
You're right though, it truly is amazing any of this works given the hardware it is running on!! |
Beta Was this translation helpful? Give feedback.
-
|
And the quality/completeness of all this is superb indeed! The documentation, the APIs, and all are all really well done. And the support here has been great for me. Best |
Beta Was this translation helpful? Give feedback.
-
|
I have the same meter. How good does it work now? Care to share your settings? And what housing did you use? I'm printing this adapter as we speak: https://www.thingiverse.com/thing:7044413 |
Beta Was this translation helpful? Give feedback.
-
|
@dbrb2 Did you also have the problem of a 3 being recognized as a 5? I think it's because of that horizontal stripe in the last number. I used to have that before, then moved the camera (a lot) closer but it's still happening. I just switched to greyscale too now, let's see how it goes... Not even using the analog part just yet, will need to move the camera slightly. 
|
Beta Was this translation helpful? Give feedback.
-
|
Ok I switched to the same model as you, let's see if that works any better ... |
Beta Was this translation helpful? Give feedback.
-
|
Nah seems even worse actually this model. That first 9 gets mistaken for an 8 oftentimes, not even sure how the model could make such error. This is all a bit disappointing, seems indeed it doesn't work to well with this Honeywell meter ... Switching back to dig-class100-0173-s2-q.tflite |
Beta Was this translation helpful? Give feedback.
-
|
well with that model (and I think also with the other one) I commonly get this problem: 
Where that last 5 is seen as a 3. Someone should train the model with a honeywell meter ... (ignore the 'rate too high' error, that's because of the problems with the previous model, i just switched back) |
Beta Was this translation helpful? Give feedback.
-
|
I've got it working well now. I'll export my config and post later
…On Mon, 8 Sept 2025, 14:03 gekkehenkie11, ***@***.***> wrote:
Nah seems even worse actually this model. That first 9 gets mistaken for
an 8 oftentimes, not even sure how the model could make such error. This is
all a bit disappointing, seems indeed it doesn't work to well with this
Honeywell meter ...
—
Reply to this email directly, view it on GitHub
<#3815 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AABFQ42VZQFSWT73HANCC6L3RVPARAVCNFSM6AAAAAB7LH2RNGVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTIMZTHEYTQMI>
.
You are receiving this because you were mentioned.Message ID:
<jomjol/AI-on-the-edge-device/repo-discussions/3815/comments/14339181@
github.com>
|
Beta Was this translation helpful? Give feedback.
-
|
Awesome! Since i'm struggling big time with this meter |
Beta Was this translation helpful? Give feedback.
-
BTW, one last question, when you're saying 'working well' what do you exactly mean? No errors at all? Or occasionally still errors? Thanks!! |
Beta Was this translation helpful? Give feedback.
-
|
I used the following config, and also this 3d mount (not designed by me) watermeter_2025-09-08_13-19-11.zip 
|
Beta Was this translation helpful? Give feedback.
-
|
Awesome!! Thanks a million, going to try your config. BTW, just curious, what 3d mount are you using, since it's not in the zip. Although I have very good results with the adaptor ring I posted and then using the 3d mount that came with this project, but without the 5 cm ring, so as to get the camera at about 4-5 cm from the meter. |
Beta Was this translation helpful? Give feedback.
-
|
Oh sorry I meant to post the link I found the height from the camera to meter wasn't quite right for me. The "right" solution would have been to edit the model. But in the end I found that a wago 221 I happened to have nearby was the perfect spacer, so flued a few of them in between the 3sp32 mount and the main body. I ended up using greyscale models for the analogue dial: Even with the colour image - as it just seemed to work a bit better 
|
Beta Was this translation helpful? Give feedback.
-
|
It's interesting that you're also using model dig-class100-0173-s2-q.tflite now, I also figured that works best with this meter. I just tried using ALL your settings but that got me an image that was almost purely white. I think it must be because you're using a black casing, mine is transparent so quite a bit of sun comes in. Which of course is not very handy, having to deal with day/night differences, I think I'll reprint it in black. |
Beta Was this translation helpful? Give feedback.
-
|
Interesting.
My print is white, but fairly opaque. But it could be there are some camera
differences
Annoyingly I just found out there is a pipe spur under my concrete floor so
my water meter doesn't actually measure all my usage (though the water
board know from their street meter), but that's not an AI problem :-)
…On Mon, 8 Sept 2025, 19:27 gekkehenkie11, ***@***.***> wrote:
It's interesting that you're also using model
dig-class100-0173-s2-q.tflite now, I also figured that works best with this
meter. I just tried using ALL your settings but that got me an image that
was almost purely white. I think it must be because you're using a black
casing, mine is transparent so quite a bit of sun comes in. Which of course
is not very handy, having to deal with day/night differences, I think I'll
reprint it in black.
—
Reply to this email directly, view it on GitHub
<#3815 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AABFQ44O4LZXI4OUUXSS44D3RWU7BAVCNFSM6AAAAAB7LH2RNGVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTIMZUGI2DQOA>
.
You are receiving this because you were mentioned.Message ID:
<jomjol/AI-on-the-edge-device/repo-discussions/3815/comments/14342488@
github.com>
|
Beta Was this translation helpful? Give feedback.
-
|
Ok I'm now using all of your setting, with only this change: CamAec = true If I set it to 'false' like you have the image becomes almost purely white, so I'm setting it to true. Everything else is the same (except for alignment and ROI's of course) The result is this, going to see how this works. At least the first recognition is going well haha: |
Beta Was this translation helpful? Give feedback.
-
|
Well enough that I can't recall the last time I noticed any problems with
its reading.
No unexpected jumps in readings for instance, and correct when I check it
against a photo.
But I should caveat that but saying I don't watch it all the time, so there
could well be occasional errors i miss.
This was the esp32 I used
https://www.amazon.co.uk/gp/aw/d/B0D8F6NNVK?psc=1&ref=ppx_pop_mob_b_asin_title
…On Mon, 8 Sept 2025, 23:28 gekkehenkie11, ***@***.***> wrote:
I've got it working well now. I'll export my config and post later
BTW, one last question, when you're saying 'working well' what do you
exactly mean? No errors at all? Or occasionally still errors? Thanks!!
—
Reply to this email directly, view it on GitHub
<#3815 (reply in thread)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AABFQ4573O5QOEVC2ULV3KT3RXRHLAVCNFSM6AAAAAB7LH2RNGVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTIMZUGQZDINQ>
.
You are receiving this because you were mentioned.Message ID:
<jomjol/AI-on-the-edge-device/repo-discussions/3815/comments/14344246@
github.com>
|
Beta Was this translation helpful? Give feedback.
-
|
Awesome!! BTW one thing I actually just noticed.. I was re-reading the ROI part again and actually different models need different ROI configuration. Since we're using dig-class100, we should do it like this according to the manual: The height of the outer rectangle should be set to the upper and lower edge of the number window So I just reconfigured my ROI's and now it looks like this: 
So, just like the manual states, the boxes now align with the number windows. Going to see if this makes any difference ... |
Beta Was this translation helpful? Give feedback.
-
|
Even with all your settings AND after my tweaks above, I'm still getting this error where the damn algo mistakes a 3 for a 5 in the last digit (the one with the horizontal stripe in it). Actually it always seems to mistake it for a 5.3 and it SEEMS that the actual value is either 3.2, or 3.3 Because this is what happens: 03999443.1,3999442.9,3999442.9,0.000000,0.0,no error,0.1,3.4,9.2,9.2,9.1,3.9,3.8,3.1 |
Beta Was this translation helpful? Give feedback.
-
|
as you can see, obviously the real number must be either 3.2 or 3.3 (because as of 3.4 it detects correctly again) and it reports that as a 5.3 and then all the next values are incorrect until it actually really reaches 5.3 |
Beta Was this translation helpful? Give feedback.
-
|
@dbrb2 if you check your latest 3 data files, there's not even a single error there at all? Because i have consistently that issue that I just descrdibed above, this morning it happened again: 4008092.9,4008092.9,4008092.9,0.400000,0.4,no error,4.1,0.1,0.1,8.0,0.1,8.9,2.9 Again the same mistake for a '5.3' and the real value in this case obviously must be between 2.9 and 3.6 (in the other example it showed to be between 3.1 and 3.4). I think this must be a problem with the model. I already was looking at training the model myself but it seems it's quite a bit of work, collecting photos and labelling them. As for now I guess I can live with this problem. Then again, maybe it would be as simple as using the original data, then deleting all 3.1, 3.2, 3,3 and 5.3 photos and just use those four photos from my own meter and train the model with that? |
Beta Was this translation helpful? Give feedback.
-
|
I have to admit I haven't looked recently. I will try to check if I get
time later
…On Wed, 10 Sept 2025, 15:39 gekkehenkie11, ***@***.***> wrote:
@dbrb2 <https://github.com/dbrb2> if you check your latest 3 data files,
there's not even a single error there at all? Because i have consistently
that issue that I just descrdibed above, this morning it happened again:
4008092.9,4008092.9,4008092.9,0.400000,0.4,no
error,4.1,0.1,0.1,8.0,0.1,8.9,2.9
4008095.3,4008095.3,4008095.3,2.400000,2.4,no
error,4.0,0.1,0.1,8.0,0.1,8.9,5.3
4008093.6,,4008095.3,,0.0,Neg. Rate - Read: - Raw: 4008093.6 - Pre:
4008095.3 ,4.1,0.1,0.1,8.0,0.1,8.9,3.6
4008093.8,,4008095.3,,0.0,Neg. Rate - Read: - Raw: 4008093.8 - Pre:
4008095.3 ,4.0,0.1,0.1,8.0,0.1,8.9,3.8
4008094.1,,4008095.3,,0.0,Neg. Rate - Read: - Raw: 4008094.1 - Pre:
4008095.3 ,4.1,0.1,0.1,8.0,0.1,8.9,4.1
4008094.4,,4008095.3,,0.0,Neg. Rate - Read: - Raw: 4008094.4 - Pre:
4008095.3 ,4.1,0.1,0.1,8.0,0.1,8.9,4.4
4008094.5,,4008095.3,,0.0,Neg. Rate - Read: - Raw: 4008094.5 - Pre:
4008095.3 ,4.1,0.1,0.1,8.0,0.1,8.9,4.5
4008094.7,,4008095.3,,0.0,Neg. Rate - Read: - Raw: 4008094.7 - Pre:
4008095.3 ,4.0,0.1,0.1,8.0,0.1,8.9,4.7
4008094.9,,4008095.3,,0.0,Neg. Rate - Read: - Raw: 4008094.9 - Pre:
4008095.3 ,4.1,0.1,0.1,8.0,0.1,8.9,4.9
4008095.1,4008095.3,4008095.3,0.000000,0.0,no
error,4.1,0.1,0.1,8.0,0.1,8.9,5.1
Again the same mistake for a '5.3' and the real value in this case
obviously must be between 2.9 and 3.6 (in the other example it showed to be
between 3.1 and 3.4).
I think this must be a problem with the model. I already was looking at
training the model myself but it seems it's quite a bit of work, collecting
photos and labelling them. As for now I guess I can live with this problem.
—
Reply to this email directly, view it on GitHub
<#3815 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AABFQ43JKZ5V3S3BGTAAPSL3SALWJAVCNFSM6AAAAAB7LH2RNGVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTIMZWGI2TAOA>
.
You are receiving this because you were mentioned.Message ID:
<jomjol/AI-on-the-edge-device/repo-discussions/3815/comments/14362508@
github.com>
|
Beta Was this translation helpful? Give feedback.
-
|
Ok, my plan is to train our V200 meter, let's see how this goes (see that training discussions here on this site) I guess it's not even necessary, since 99% of the times recognition goes well. But it's a nerd thing of course to see if we can get to 100% :) |
Beta Was this translation helpful? Give feedback.
-
|
BTW when saving data for training, I found the 'problematic' 3.2-3.3. It recognizes this currently as a 5.4: |
Beta Was this translation helpful? Give feedback.
-
|
Here's the (digits) model i've trained with my meter: https://github.com/gekkehenkie11/android_device_samsung_trltexx/blob/cm-12.0/dig-class100-s2_HONEYWELL.tflite Next, might look into the analog part as well ... Only problem is that I need to either move the camera a bit to the right as the analog meter currently doesn't fit onto the photo as you can see above. So I'd need to 3d print a new lid with a new hole location, or need to move the cam slightly up. I already looked into moving the camera hole a bit to the right on how to do that, but damn, that's not easy!! Not even sure why it's so complicated in the era of AI, should be super simple to select the camera hole and just move it, but it's not.. So will think about how to proceed here |
Beta Was this translation helpful? Give feedback.
-
|
https://jomjol.github.io/neural-network-digital-counter-readout/ |
Beta Was this translation helpful? Give feedback.
-
|
Thanks - I'll try out your model this evening! |
Beta Was this translation helpful? Give feedback.
-
|
I've now ran the meter for 12 hours with the new model, a little early maybe to draw conclusions but I haven't seen the problem anymore in the last red digit, where I'd get that several times a day in the past (mostly errors with 3's being recognized as 5's). But will keep on monitoring. I did see an error in the black digits though where it would mistake a '1' for a '7'. This happened occasionaly before with the other models too. I think this wasnt fixed because the black digits don't turn that much in 3 days of data of course (actually it weren't even 2.5 days). Maybe it's an idea to change the setting to take a photo every 10 minutes and then, say, for 10 days and train again with that? Although this seems only a minor issue, it happens 2 times in the last 12 hours. I will keep monitoring |
Beta Was this translation helpful? Give feedback.
-
|
BTW I reverted to my old camera settings, i feel they work better for my cam. I'm pretty sure you have a different OV2460 than I have. There are several manufacturers out there. 
Here you can also see that black '1' that sometimes gets mistaken as a '7' |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
I have a V200 meter that seems to be struggling with the analogue dial used for the fine readouts - and was wondering of anyone could make some suggestions.
I have been trying out vartious models for both the analogue and digital readouts without a great deal of success so far:
Beta Was this translation helpful? Give feedback.
All reactions