Image Search Breakthrough: 0.98+ Precise Images of 500,000+ topics | Hacker News Reader

Image Search Breakthrough: 0.98+ Precise Images of 500,000+ topics

rdf.ookaboo.com

77 pointsPaulHoule17 comments

In TechCrunch-style blogs, every post has a large image at top (often just a stock photo) that's also used to represent the post as a thumbnail in post lists.

You could automate providing these images for new posts by mashing up this annotated database of CC-licensed images with the OpenCalais API. OpenCalais analyzes text to extract named entities (companies, technologies, people, even social tag recommendations) -- so match those terms with the image database til you get a hit and you've got an image with no work done by the author.

Maybe something I'll play with later tonight.

I wrote a PHP wrapper for the OpenCalais API some years ago. https://github.com/dangrossman/PHP-OpenCalais

For an example of how this is done with google images check out this project http://logolifter.com/

Am I missing something or is this site specifically designed to enable copyright infringement? I don't think the little disclaimer saves it. Contributory infringement is illegal in the US, and this tool lifts images from Google Image Search, 99.9% of which are not licensed for distribution. If this is your site, I'd consider replacing the image source with something like Flickr whose API which supports search-by-license.

Google Image Search indexes Ookaboo, not the other way around.

All of the images were published under a public domain or creative commons image elsewhere; there is a small probability that this information is incorrect, in which case we remove images as per DMCA

Hi Paul. I was referring to LogoLifter, the comment I replied to, not Ookaboo. LogoLifter's images are from Google Image Search, not from Ookaboo.

Or you could use Zemanta: http://www.zemanta.com/

Please fix the title ("Image Search Breakthrough: 0.98+ Precise Images of 500,000+ topics").

Unless I'm misreading this radically, this is a learning dataset of 1,000,000 images, labelled with a precision of 98%.

This isn't a breakthough in image search. It is helpful for anyone working on image search, but appears to have been manually labelled.

You are right, the title is grossly misleading, but I don't think this was manually labelled. They have this to say about the association of pictures and topics:

Ookaboo attains high precision within it's point-of-view by having a very permissive definition of "depicts," which means roughly "this image could appear in an encyclopedia entry for..." For example, in the case of dbpedia:Pablo_Picasso, we'd accept a portrait of the artist, images of his works, or a house that he lived in.

This leads me to believe that the topic associations were derived automatically based on information found on the page in which the picture was embedded.

Yes.

To be clear, I suspect a large number of the images come from Wikipedia/Wikimedia, and are manually categorised there. This software scraped the images and the categories.

So is this meant for actually getting images or as a learning set for ML?

Related: you can grab 100k images files with labels[1] from the old ESP game[2].

[1] http://www.cs.cmu.edu/~biglou/resources/ [2] http://en.wikipedia.org/wiki/ESP_Game

This looks quite impressive and useful. Could someone explain what the precision metric refers to? The documentation states:

But that doesn't really explain how the number 0.98 was arrived at.

EDIT: made this into a straightforward question rather than speculation.

1000000 * 0.02 = 20k

This implies there are 20k wrong images in the dataset, sampling can find some of them

doesn't it mean could be 20k wrong images?

Yes, you are right. 20k is an approximate number.

It's hard to give a number for Ookaboo's precision because of the difficulty of evaluation. Different raters disagree with each other around a rate of 0.02.

This looks very useful. One of my customers needs labeled images that can be used with attribution. Thanks for posting this.

Reply on news.ycombinator.com