Image may be NSFW.
Clik here to view.
Text input search is passe. These days, it’s all about voice and image recognition, as seen in Apple and Google battle against one another.
While the two tech giants are duking it out in open warfare, a Singapore startup called ViSenze — a spin-off from an NUS research project — has entered the search arena via a backdoor. It has partnered with e-commerce players Rakuten and Clozette to deploy its technology enabling shoppers to upload an image and receive search results consisting of fashion items with matching clothing types, style, colour, and patterns.
The Fashion Finder can be used on two sites: www.oshare.com.tw, a joint venture between Rakuten and Singapore fashion community Clozette, and www.rakuten.com.tw/event/funsearch.
According to ViSenze, the technology has many other applications. Cars and consumer electronics websites could find it useful, while advertisers can use it to match a search result page with a visually-resonant ad. For example, a user searching for shades could see an ad showing sunglasses of a similar make, model, and color.
I find the technology reminiscent of Google’s very own image search feature. Another startup that is knee-deep in the image recognition space is Graymatics, which is applying the technology to advertising and filtering.
Relevancy is the name of the game. Whether or not ViSenze will succeed depends on its ability to serve up useful search results.
Measuring user intent is a tricky business however, since some shoppers may know exactly what they want while others adopt a search-and-see-what-happens approach (I suspect women may tend towards the latter).
In any case, I decided to put the technology to the test.
The first thing I noticed is the lack of drag-and-drop functionality similar to what Google employs in image search and Gmail. At the moment, Fashion Finder only enables URL search and image upload, but these methods are cumbersome.
I then tested the search engine by using images from the Internet. Here are the results (search input has a red border):
Search #1: It’s an admittedly complicated photo, with the background and foreground in focus. But how else to test the prowess of ViSenze’s technology? Besides, users may want to upload Outfit of the Day (OOTD)-type photos, so the tech should ideally be competent enough to decode such pictures. Looking at the results, the algorithms were able to detect a bag in the photo, but the search results were rather inconsistent.
Image may be NSFW.
Clik here to view.
Search #2: To make things easier, I used an image with a blurred background that puts the subject into focus. But for some strange reason, the search engine gave me stripes and more stripes. Maybe it has something to do with the bands of shadows in the image I used?
Image may be NSFW.
Clik here to view.
Search #3: Since the background could be a distraction, I trield a close up. The search engine recognized the two-layer ensemble of vest and T-shirt, although the search results was a mish-mash of different styles. It didn’t even recommend a similar vest.
Image may be NSFW.
Clik here to view.
Search #4: Let’s pare down to something simpler. This time, the search engine is able to detect the color of the image and make some interesting recommendations. But what’s with the female clothing?
Image may be NSFW.
Clik here to view.
Search #5: I stripped out the fluff. Plain white background with a simple blouse. This time, Fashion Finder is less confused, nailing the colors but still presenting divergent styles.
Image may be NSFW.
Clik here to view.
Search #6: Hot pink handbag throws up more recommendations for hot pink handbags. This is a no-brainer.
Image may be NSFW.
Clik here to view.
Search #7: It did a good job of recognizing strips and the type of shirt. The female blouse in the top right was the only exception.
Image may be NSFW.
Clik here to view.
The limitations of the tech became apparent after a dozen searches. Complex OOTD photos featuring an ensemble faze the search engine. Instead, uncomplicated, high-contrast images focusing on single item types seem to work best.
For now, the search engine is unable to account for a range of user behavior. Perhaps having toggles that allow for ‘discovery’ searches versus ‘specific’ searches might help. Having the ability to prioritize or filter results by color, texture, make, and even gender might be useful too.
Anyway, with further user feedback in the coming days, we should see the technology improve. For all we know, users might not really care about relevancy at all.
But in the meantime, Rakuten or VizSense might want to consider creating a guide to teach shoppers how to circumvent the tech’s limitations.
The post ViSenze’s visual search tech on Rakuten Taiwan has promise, but can’t handle OOTD images appeared first on SGE.