I am trying to use machine learning / neural networks to count how many units of an object type on front row of a shelf in a photo (let's call that object as "green beer").

As you can see in the photo below, when captured at direct angle, the green beer cans on behind row and front row are almost indistinguishable, and normal object detection neural networks can easily be confused.

Note: Beverage brand is intentionally blured.

enter image description here

I have an idea that if somehow I can combine photos with different angles and build a 3D space based on them, then I can count more precisely how many beer cans on front row only.

enter image description here

I am really really new to machine learning, so please suggest me some algorithm names or links if someone has already encountered this situation. I prefer neural networks solution but any other kinds of machine learning are also welcomed.

Thank you very much.

  • $\begingroup$ If you are "really really new to machine learning", I would recommend starting with simple problems, such as using neural networks for MNIST classification, until you feel like "advanced in machine learning". By then I am sure you will know names of the algorithms you are looking for. What you are trying to achieve seems quite difficult, i.e. not a beginner level problem. $\endgroup$ – Jan Kukacka Feb 26 '18 at 12:47
  • $\begingroup$ This is my work project, and since I am working in R&D department so I am the first to study and experiment with new technology. And it's not easy to find ML experts with affordable salary in our country. Although I'm new, I'm very willing to learn. $\endgroup$ – hirikarate Feb 26 '18 at 13:10
  • $\begingroup$ I understand your position. But as an expert in an R&D dept you should also be able to tell: Solving this problem is currently above our capabilities. On the other hand, if you are dedicated to learn, I just provided a pointer where you could start learning and understanding the technology. $\endgroup$ – Jan Kukacka Feb 26 '18 at 13:13
  • 1
    $\begingroup$ @hirikarate This problem you're asking is non-trivial. To "detect object in 3D", you have to tell the spatial location, or at least the relative distance betwn objects. But to do that you need to segment the image into parts. Figure out how to do these tasks first. $\endgroup$ – horaceT Feb 26 '18 at 14:05

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.