Image recognition with a single reference picture

asked 2019-08-07 10:12:05 -0500

Hey all! After building my app using Vuforia (great stuff, until you have to pay haha), i was wondering if it is possible to do picture (not facial or object) recognition without a dataset. I know my way around basic programming and eager enough to dive in a bit deeper, but I would like to have some pointers to start.

Looking at the data that Vuforia generates of a single image (multiple points of recognition based on light spots to my untrained eye), how would one do the same using openCV?

Cheers!

edit retag flag offensive close merge delete

Comments

1

it's called "one-shot" learning, an active research topic.

berak gravatar imageberak ( 2019-08-07 10:14:49 -0500 )edit

Thanks, ill dive into it after diner!

Since I am allowed to outsource this part of the project, is there anyone here able and willing to create a working system (no need to be perfect, but a decent foundation with sufficient documentation is a must) for our server?

TheIndieStation gravatar imageTheIndieStation ( 2019-08-07 11:43:08 -0500 )edit

berak gave the right keyword - for your problem you can check for facenet. I even have a facenet micrososervice in python with a custom face detector (the one used by facenet sucks). So if your are willing to pay - i can consult /develop with you. But i am sure you can also do by yourself!

holger gravatar imageholger ( 2019-08-07 15:14:30 -0500 )edit

The thing is we are on a time shedule. Are you open for a chat tomorrow 12:30 GMT+1?

TheIndieStation gravatar imageTheIndieStation ( 2019-08-07 15:26:28 -0500 )edit

@holger. Can you send me an email to (removed)? I would like to discuss this project with you now I have green light on a budget.

TheIndieStation gravatar imageTheIndieStation ( 2019-08-12 11:36:13 -0500 )edit

Just to be perfectly clear, coming from someone doing active research in this field, single shot and one shot learning is far from solved. On challenging datasets getting an average precision of about 55% is already considered a breakthrough. Just to give you an impression of how difficult this will be :D

StevenPuttemans gravatar imageStevenPuttemans ( 2019-08-20 08:07:15 -0500 )edit