Can Google Solve the Image Search Problem?
The New York Times is out with a piece today on the latest paper put out by Google on image search. The paper was presented at the International World Wide Web Conference last week in Beijing and purports to use something called “VisualRank,” a computer algorithm, to combine image-recognition software with rank and weighting.
According to the Times, Google has a team of “150 Google employees” working on a new scoring system of image relevance. Google claims that this “team” returned 83 percent less irrelevant images in a sample group of 2000 popular product inquiry search terms.
Last year Robert Scoble and I had the opportunity to meet with Marc Levoy, a photography researcher down at Stanford, who said that he was working on image search with Google. I tried to get Marc to elaborate on what Google was working on or how they were approaching image search but he kept his cards close to his chest and would not elaborate specifically. I’m assuming that some of the work that Marc was doing may have to do with this project.
Although Google has never been exactly clear on what they are doing with image search, it is something that I’ve been watching over the years.
Previous to this paper, the most significant insight we had on what Google was up to had to do with their “Image Labeler game.” With Image Labeler Google allows people to compete against each other to label certain images on Google for time. One person wins and one person loses. I’ve played the Image Labeler game before and have been critical of it. It’s about as much fun as watching paint dry.
I’m not sure how successful this game has been for Google, but I doubt that it’s significantly enhanced their image search. And so now Google returns to the ever elusive artificial intelligence angle. Google has always hated having to rely on actual humans vs. algorithms to rank and rate images so this comes as no surprise to me.
Remember a few years back when there was a big rumor out that Google was going to acquire Riya? Riya claimed to have technology that could effectively recognize faces in your photographs and auto tag people in photos based on this facial recognition software.
I tried Riya’s technology shortly after they launched it. Uploaded thousands of photos to their site, spent hours training their software on faces and in the end their results were abysmal. Despite tons of dough being dropped on Riya by some pretty naive VCs, their software quite simply did not work.
I’d heard rumors that some of Riya’s work wasn’t even computer algorithm. That they were having people in India actually physically go through the photos to try and correctly label them. Apparently people in India had a hard time telling one white person from another. These are just rumors.
Personally, and despite the hype that you hear from time to time from the likes of Google and Riya, I seriously doubt that you will see image recognition software good enough to effectively enhance image search anytime soon.
Tagcow came out about a month ago offering a similar mysterious service that would categorize your images with tags. They didn’t say *how* they tagged your photos, but TechCrunch later reported exactly how they were doing it. They were paying people $1.20 per hour with Amazon’s Mechanical Turk to have humans do this work.
The problem with image search for a company like Google is that there is no business justification for spending $1.20 per hour to have someone tag or rank images. Google doesn’t even put advertisements on their image search pages and I’d bet that ads are highly ineffective here.
Google does image search because they have to. It’s a me too sort of thing more than anything and I can’t see them spending money on humans to do the work as it would run contrary to their entire philosophy of leveraging a computer to do a human’s work. I was in fact surprised even when I first heard about the Image Labeler game — but Image Labeler relies on free, volunteered labor. Free, volunteered labor by who? I have no idea. I simply cannot imagine *anyone* being bored enough on the internet to want to tag images for free for Google.
The king of image search today of course is Flickr. Yahoo smartly realized early on that as valuable as Flickr was as a fun photosharing playground for people, it also had an ancillary benefit (in fact maybe even a primary reason — remember folks, the search team at Yahoo bought Flickr at Yahoo not the photo team) of providing kick ass image search.
Like Image Labeler, Yahoo relies on free labor for their organization and rank of photos as well. Except, in Flickr’s case it actually *is* fun. It’s part of the Flickr experience and when users tag their photos for their own purposes and for search purposes and then other users fave, comment, view (read image rank) these images, you tend to get superior image search results.
This will likely benefit Microsoft most of all — assuming a Microsoft takeover takes place at some point.