@deepmike, Interesting project. I have seen something similar on the web, namely here, https://github.com/paucarre/tiefvision.

Would like to collaborate. Send me a message on hackernews, if you feel likewise.

I did see tiefvision, yes. Somehow, their approach didn't quite work for me, although that was the first attempt. Maybe it's good when you have only few classes of images ("only dresses") and modest amounts of training data.

..there are no direct private messages on HN, right? I have an email in the blog post, just in case there is no other way

How are you planning to pay for webscale computational resources?..

Good question. Right now I am using AWS spot instances + credits bought for Bitcoin. This makes the thing maybe 70 times cheaper than on-demand instances paid normally.

But when I go above several billions of images, I'll need some investment for sure :-)

Tineye indexed about 20 billions of images, and my estimates show that I should be fine with my own resources up to 1B. This is good enough to improve and test the tech

what will this do that google / baidu search dont do

Google searches for exact match + image classification + text search, and also recognizes some famous images (such as Eiffel tower). E.g. if Google sees an unfamiliar image, it converts it to something like a "red swivel chair", and searches for that (so no way it can find the same object unless it is really unique).

I want to make it possible to just snap a picture of anything and find this object. Try to take a picture and search for it on Google Images.

Although, I should say, image search in Baidu works way better than Google's

How do you suppose the program should perform image recognition better than existing tools? Not a lot of experience in this realm, just curious