Hacker News Reader

Comment by lucaspiller | Hacker News Reader

Very nice! What are you using to search the photos by colour and feature?

Probably performed dominant colour analysis on each (like color-thief), sort by similar ranked colours using a closeness transformation like LAB? No idea for the feature though.

Is there an algorithm for what you just described? I'm currently researching for color classification and it seems quite a complicated issue.

Hi, If you want you can contact me. I created a powerful algorithm to detect dominant colors in an image using K-means clustering and lab color space.

Examples: https://twitter.com/kamy22/status/479040852028051456 https://twitter.com/kamy22/status/472517258418606080

Have a good day!

What about feature? If I want a dog, how do you determine which images contain a dog?

You have to do a lot of research to solve these type of problems. I think that neural networks and machine learning are the best way... but it's a complex problem.

Here you can find awesome publications (http://rodrigob.github.io/are_we_there_yet/build/classificat...). It's something like a bible of neural networks :P

Considering that companies like Google are so good at this, why build your own photo site? Why not upload all the CC0 images to a public Google Photos library?

I like your question. Ok, you can use 500px, Flickr, Google and other sites... But... In my personal opinion, a developer should be curious. I'm a dev and, for this reason, I like to give myself a challenge. It's a good way to learn a lot, to discover new solutions, to meet new people, to improve my skills, to create something new. So... you can use Google Photos library or you can consider to create something different (because definitely your solution will be different from the others). It's a choice ;)

I'm currently trying to do this for my own side project, my current plan was analyse each photo using one of the many color analyser algos out there (color-thief seems to be popular/efficient), but the problem is, color is a very complicated thing to compare. I think euclidean distance for two LaB colors is the best for complexity to result ratio (and RBG to LaB is an easy/well-solved problem).

I've come across this: http://developers.lyst.com/2014/02/22/color-detection/

Seems very good but I'm still researching. Hope it might help you.

FWIW digikam has a colour search - could check their code and see what they're doing.

something like: http://freecode.com/projects/mactorii should work

Reply on news.ycombinator.com