How to count the number of objects of the same type in a photo?

Question

Neural networks

There is a task (for me this is an academic interest) to count the number of objects of the same type in the photo. Photo example .
Conditions:

• Different weather/lighting/angle
  
• We take pictures on the phone and send to the bot in telegram
  
• Counting time - at least less than 10 seconds
  
• Ideally, perform calculations on a hosting, without involving a GPU
  

At the output, it is desirable to get the same photo marked with bounding boxes with their numbering and total number (for visual control). Example .
I tried to adapt the code from the article for my task - it turns out badly , most likely due to the "field" shooting conditions. The rolls sometimes lie not exactly in the plane and/or are slightly covered by something. I already foresee false positives for reflections in a puddle or some kind of car wheels.
Another example came across on cats . There we are talking about some pre-trained classifiers and it is unlikely that there will be rolls among them))
Therefore, I began to move towards neural networks. I found such dissertation - Automated counting of objects in the image.... A difficult topic for me.

1. Is it still possible to get by with "ordinary" algorithms and not get into machine learning?
   
2. If yes, what would you recommend. Suddenly, I just lost sight of the ready-made solution for my task (from the same OpenCV)
   
3. If ML needs to be studied, then I ask for advice on a dissertation. How realistic is it to repeat the experience of researchers.
   Or maybe there is a ready-made neuron, which is enough to feed your dataset.
   

Answer 1

I'm working on a similar problem. The author contact me (perhaps through the answer here, we will decide there), I will need training pictures, maybe I can file my network for your problem, but I do not promise anything in time.

Answer 2

YOU DO NOT HAVE A SPECIFIC QUESTION.
And here he has nothing to do.
Read this for a start
https://habrahabr.ru/company/ods/blog/337548/

Answer 3

www.pyimagesearch.com/2015/11/02/watershed-opencv
but in general you have to figure it out yourself, including pattern matchong

Answer 4

1. Compile a dataset from several ends of these rolls. Those. in fact, this is a 6-tihedron with black triangles at the edges (0 - no gap, 1 - gap):

001100
110011
000000
110011
001100

2. Then, determine all similar parts in the image from the trained HC vector (detection of the end of the roll).