Detection of playing cards - Methodology?

Python
1

Hello,

I’m developing a solution that detects playing cards in a given real-world scene (subject to perspective and euclidean transformations) and the end result hinges on properly estimating the bounding box of any given card at any given angle.

Due to the symmetry and lack of features of the cards in use, it’s become apparent that using feature extraction alone will not provide a robust enough solution as it very easily falls victim to matching keypoints to the mirror side (of either the cards as a whole or any glyph on the card), resulting in incorrect homography (as seen in the attached pictures)

Before I attempt to do this on my own, I would like to ask if any of the more seasoned OpenCV developers have an idea of how I could pre-process the image and run detection in a way that is more conducive to obtaining good results. (no code is necessary, just the process in case you worked on something similar or otherwise have a hunch)

Many thanks

res_1
res_11280×558 36.4 KB

res_2
res_21280×558 37.9 KB

res_3
res_31280×558 35 KB

2

the features on that card form a line, not a plane. see how the letters and symbols are on a line?

3

Indeed. Is there any way I can exploit that?

4

the artwork on the card is insufficient. you need its outline as well.

locating and identifying cards is not a novel problem, meaning there’s been decades of people coming up with ideas and working on them. you should do literature review.

5

I will. Many thanks!

6

You could look at the different apps that read magic the gathering pokemon yugioh (neuron) perhaps to give you a clue as to how they do it, also the genteman who programmed Delver Lens years ago may also have some insights. Also there is another prevalent considerations: glare. Alot of card readers have great difficulty when there is any light glare on our cards.