Hacker News Reader

Comment by rocha | Hacker News Reader

Another option is an image recognition system trained on already existing cards. No need for special ink or qr code, just a camera as the reader.

View on HN

Just OCR against the card-corners may be sufficient.

I guess you also need suits, but that's only four special characters (all of which may be in more-modern OCR-libraries).

As I said in another comment, normal cards are easy. The issue is with everything else where each card might have a picture with different animals and you are looking for the lama giving you the bird.

I think it is very doable for card games with thousands of different cards like MTG. I remember there was an old school SIFT/SURF image retrieval system that was able to deal with tens of thousands of album covers super fast. That was more than 10 years ago IIRC.

If the manufacturer is willing to provide the entire db of cards or a well-trained model that could be embedded or a ready mobile app, that would be great. However, waiting on them is not a winning strategy. The other option is community contributions, but you need a large community. This will be welcomed as well, but again not something that should be relied on. Both developments would be wonderful to happen, however in the meantime I'm looking for something cheap, fast, and easy to implement that I can do on my own in the local scope of my deck.

For some of the card games I am familiar with (Netrunner, MTG) there are resources online (wikis, dbs, deck builders) that have all the graphic assets. Alternatively, you can scan or photograph each card in your deck, but that may be time consuming.

As for software, there are engines out there that do content-based image retrieval (CBIR) out of the box [1]. It should be possible to build something quickly in OpenCV. You may be able to get away using simple image template matching by putting a few constraints on how the camera sees each card. Something more robust can be also be build using image descriptors and approximate nearest neighbors, as in [2].

[1] https://en.wikipedia.org/wiki/List_of_CBIR_engines

[2] https://blog.francium.tech/feature-detection-and-matching-wi...

Reply on news.ycombinator.com