The raging social media debate on a simple bodycon dress has polarized the Internet. One poll had 69% of respondents seeing white and gold, with the rest seeing mostly black and blue.
Theories emerged quickly, including one that attributed the difference to left-brained versus right-brained people. Really?
Even the eye doctors who weighed in explained that it is not unnatural for different people to see different colors because of the way the human brain perceives visual signals.
Well, whatever is, that is why the visual system and visual perception by the human brain remains so fascinating, isn’t it?
With that much fun happening, we (image recognition scientists) couldn’t resist either. We (ViSenze) decided to test our image recognition algorithms, and see what our algorithms see. After all, algorithms don’t have retinas with pigments that fire neural signals to the brain.
So technically, they shouldn’t suffer the same perceptional problems the human brain do.
But let’s state facts first. The bodycon dress is black and blue (confirmed by Roman Originals! They also confirmed they don’t yet have the white and gold version!).
So with this as the ground truth, we set out by using the same exact image (in fact, we chose from 3 different media sites – Mashable, Forbes, The Independent – to avoid selection bias) and ran them against our image recognition algorithms.
Within milliseconds, our algorithms processed the pixels of the image feeds directly, and spitted out the results.
This was what our image algorithms recognised (see below); a specific and granular set of colors picked up, including background colors.
[The color codes represent different color types in our color spectrum taxonomy. Ignore the background color identification]
The results of the same dress image sourced from the 3 different sites were all consistent, with very little differences.
Well, that’s good. But that is still quite a wide range of colors to consider. So we asked ourselves; if algorithms were people, and if they had to choose the most matching colors from a range of outfits, what would they choose as their first choice?
To avoid bias again, we specifically ran the test publicly on Clozette Shoppe. Clozette is one of ViSenze’s clients that use our visual search algorithms to search against an online fashion database. Their database has over 2 million items.
To be technically accurate, our visual search engine uses a “nearest neighbor” set of algorithms that first recognizes the pixels in the image, extracts a rich set of visual signals (eg. color, patterns, shape etc.) to conduct a search for most visually similar candidate from a database.
For this purpose, we took the first search candidate returned by the algorithm as the human choice equivalent, although our visual search engine goes on return other similar outfits as part of its recall capabilities.
Guess what it returned as first result? The same color as the original blue and black bodycon dress!
We tried it several times with the same image from other sources just to be sure. Well, our algorithms were pretty sure! (Notice how closely it also picked up a white and black dress as the next most matching candidate?)
[Screenshot from Clozette Shoppe that uses ViSenze’s award-winning visual search engine that finds visually similar fashion outfits without keywords. Caveat: the dress is a closest match on Clozette, it’s not supposed to be exact since Clozette do not carry Roman Originals.]
Which brings us to the bigger point in this debate: Can science and technology help humans?
Definitely! The human eye sees what it sees and the human brain interprets (or in this case, tries very hard to guess the colors). Image recognition algorithms don’t suffer the same problems humans do. They don’t have retinas or left-right brain bias.
If we train algorithms well and objectively, they will be precise and they will consistent. And here’s the fun part (or lack thereof), algorithms don’t argue back or engage in social media debates! They just show you what they are scientifically trained to do.
And all these while, we humans think that technology and machines are the more complicated ones. Irony, isn’t it?
(For the record, this author sees a white and gold dress!)