I didn’t see an explicit category that made sense for this so let me know if there’s a different place to put this…
I am doing a project to detect “symbols” or “glyphs” (black and white) drawn on a piece of paper. I’m looking for a set of 10-15 glyphs that can be easily detected and differentiated by a CV model.
Been trying various combinations of glyphs and want to know - what does CV find easy to differentiate at small sizes? texture? shape? filling? shading? homology (number of “holes” in the glyph)?
(or the opposite question - what types of glyphs are easy for CV confuse for another? For example, maybe the striped filling is too similar…)
The first layers learn features like lines, corners, and colors, the next layers combine those together into shapes and textures, the next layers combine those into parts of objects, the next layers combine those into objects. And all of these are learned from scratch based on millions and millions of iterations of strategic trial and error (back-propagation).
So given “enough” consistent data your model should be able to learn to detect anything from drawings to animals to oil spills to cancer.
What’s “enough” is a function of the variation in the objects and how “similar” they look under various conditions. The problem you’re describing is probably less about what your model can learn and more about what humans can most consistently draw and which objects look most different from each other in a wide variety of orientations and lighting.