Classification via one image alone

Part 1 (2019)
1

Hi all,

Here is my problem. Say I want to make a model that can classify all 52 cards in a deck. From only one specific deck. And no other decks exist that relate to it. What would be the best way to start looking into this? As the train and validation would be the exact same.

Thanks!!

Edit: one thought would be to ‘crappify’ to some degree the cards to make copies. And turn those crappified cards into the train, and the real cards into the validation, as now we have multiple images with the same label

2

Why would the train and validation be the exact same? If you do segmentation and transformation to each of those segmentations you can easily get 400+ images. I’m still wondering if that is enough though.

1 Like
3

They would be the same because there is only one that exists. Eg I don’t have multiple images of the same thing. I just have one available to me.

4

What if you do the crappification along with random crops and perspective augmentation? Perhaps to make up for the lack of data you can train a siamese net on it? Since this way you can make use of pairs of images that are of different cards too.

2 Likes
5

Thanks for the suggestions guys! I’m hesistant on Siamese as the end goal is to help a vision-impaired individual play a specific game and locate certain tokens via an app they hover above the ground. Would a Siamese or segmentation be good here? (I do not know much on Siamese and will read that article shortly).