How to properly teach to recognize objects (imageai)?

Question

Python

Hello!
There are a few machine learning (beginner-beginner) questions.
I use Python + ImageAI.
from imageai.Detection import ObjectDetection
The task, to put it simply, is to get the number of cubes in the box. All cubes are the same color, one can overlap another.
1) Is it possible to achieve normal recognition (with high accuracy)?
2) And what needs to be done for this - that is, what pictures are better to shove into training?
The principle itself is not clear to me, how the system remembers certain objects. I prefer a lot of shots, where there is one object in the frame, but at a certain angle? Or is it better when there are a lot of objects in the frame, in particular, some overlap each other?
3) How does the system react when I frame (via LabelImg) a rotated object that partially overlaps another? Should I circle all the cubes in the picture, or is it optional?

4) and finally, the simplest, probably - what is the LOSS function? No, I understand that this is the share of losses, mistakes. But, for example, if at the first iteration LOSS = 150, and at the last iteration LOSS = 15, then 150 is from what? From 1000? That is, what number reflects 100% loss?
What's the point. Now I have a set of 130 pictures. The system guesses about 80% of objects, LOSS 16. If I increase it to 200, will the quality of recognition increase? And if you increase it to 200 (or even higher) - I just want to understand which pictures to select and mark up so that the training is most effective.

Answer 1

Take a course like cs231n or deeplearning.ai. You still have porridge in your head, but the task is actually not the easiest.