Advice for detecting gas cosumption

[rradic (Robert)]

Hi Fast AI community,

I’ve started Part I(2018) course 2 weeks ago, watched few lessons, tried lesson examples and now I’m trying to train my own model for recognize gas cosumption based on gas counter images like this one.

There is a few different types of these gas counters and all of them have different position of the exact counter part(bottom bounding box from image) and each type of gas counter has it’s own manufacturer number (upper right bounding box). I suppose that I’ll have to do some combination of classification and object detection, but that’s just my assumption.

Do you have some advice for which approach should I use for solving this problem?

[hofyitulti (Hofy Smith)]

Did that really work out, for me lag is making the mess.

Regards,
H. Smith https://tutuappx.com/
https://vidmate.onl/

[rradic (Robert)]

Yes, we made it by using Darknet/YoloV2 arhitecture. But not within fast.ai

We have three models. First one for manufacturer number and gas counter localization, second one for detection of manufacturer number digits and third one for detection of gas counter digits.

It’s working quite good actually, since all of our models are trained on same dataset with 900 images. Further plan is to implement this inside fast.ai.