I’m trying to classify playing cards by color (from the game of SET). The code is here:
I have 56 training samples at the moment, which I know isn’t great, but sure seems like it should be enough to solve a simple “one of three” classes problem. Which of my assumptions or techniques is so wildly wrong?
As I see it, you need to make sure that the NN does not learn shapes but colors. To encourage this behavior, set max_zoom to 1.3 and retain default max_warp and also add vert flip and max_rotation = say 90.0. Let me know if my suggestion makes sense.
It makes sense, I’ll give it a try. I assume the theory here is that the less “wasted white space” the better. I think my OpenCV prep work is also super bad, but that’s probably a problem for after the NN is working. I assume that extracting the 12-15 cards from a full image is beyond the capabilities of a normal CNN.
Nope, or at least not simply.