Everypixel launched an AI for estimating the attractiveness of Instagram photos
Now we are going to tell you how it will make your services more efficient.
Only in the recent 24 hours, people have uploaded more than 5 billion images to the social media. Obviously, it is not the end: the pace keeps growing! To cope with the curation and moderation of all user-generated content manually is no longer possible, it requires a lot of resources and time. It is much wiser and more efficient to use artificial intelligence for this purpose, especially since we have already developed it.
How It All Started
In 2016, our team took up developing the stock images search engine Everypixel.com. Then was created the Everypixel Aesthetics, a neural network to automatically clean search results from obsolete images. We trained it with a selection of the most trendy stock photos. We didn’t want the target audience of designers to have any doubts about its work, so we created a special demo page. Everyone was able to check there how the Aesthetics works by uploading examples of stock photos and realizing how awesome are they.
But our neural network became popular far beyond the professional community. Users started uploading selfies, drawings, and photos of cats to the demo page. The AI trained in stock images did not always evaluate such content correctly. Sometimes, really bad shots were rated as 100% of coolness, when really good photos taken by a smartphone got no more than 10–20%.
So it occurred to us that it was time to train a new AI to score user-taken photos. It was what ordinary people wanted, and it was requested by developers of online stores and applications who already faced the need to filter out huge volumes of user-generated content.
We collected 347,000 photos from Instagram and asked a team of 10 photographers from different parts of the world to distribute them into 5 classes: very bad (0–20), bad (20–40), good (40–60), very good (60–80) and excellent (80–100). The data obtained were used for training the AI. Finally, the neural network began to make objective and accurate estimates of not only mobile photos, but also images taken by a professional camera.
The new neural network has become a part of the Everypixel API. It is called User Generated Content (UGC) Photo Scoring. Similarly, the already mentioned Aesthetics was called the Stock Photo Scoring.
Why Is It Better Than Human?
First, it is extremely accurate! When training this neural network, we have foreseen possible difficulties and have taken into account the peculiarities of different areas of user photography. So it can’t be misled by texts typed over a photo and by flatlays from online stores.
Secondly, it is quite objective! The main criteria for estimating images from the training sample were only the technical parameters: clarity, composition, exposure, framing, and so on. For example, a funny, but a blurry photo of a cat still received a low score. While high-quality portraits always received the highest score — even though model’s appearance is far from generally accepted standards of beauty.
Finally, our neural network is very hardworking and does not ask for vacation. Moreover, it’s very cheap — 1,000 queries will cost only $2. Just compare it with the human moderator salary. For those who need huge volumes of queries, we are ready to develop an Enterprise plan.
Where to Use
How can UGC Photo Scoring help your service? Let’s find it out using some examples.
One of the possible application scenarios is to rank search results by photo quality. The girl has uploaded her blouse into the eBay, but the photo makes it impossible to see what design it is. It is unlikely that such photos will attract potential buyers; the product can be lowered in the search results. At the same time, the girl can be advised to choose a better photo.
As for recommendations… How to choose the best photo from a series of almost identical shots is a typical puzzle for social network users, isn’t it? The same can be said about any services where people need to choose the best shot to attract attention: from Instagram to Tinder. Now, UGC Photo Scoring can help to cope with it
If you are developing a mobile photo editor, why not to warn the user that he has overdone with filters. All you have to do is to offer a simple photo check when the user tries to save it. If the photo gained more than 30%, according to UGC Photo Scoring, it’s cool! If the score is not so high, you can warn the user that there is still something to work on.
And we have got a lot of other examples. Want to assess how well this technology will work in your specific case? All users of the Everypixel API have 3,000 free queries per month. Just sign up.
Anyone can get an impartial assessment of his photography talent at once by clicking the UGC Photography Scoring tab on demo page of the service. Try it right now! :)
The Everypixel team promises that our neural network will soon expand its functionality: it will not only “criticize”, but also automatically improve photos. Don’t want to miss the update? Then subscribe to the Everypixel official Facebook page.
Here are some examples proving that Stock Photo Scoring and UGC Photo Scoring estimate images in a different way.
And finally, let’s compare two neural networks once again:
Stock Photo Scoring was trained on stock photos;
UGC Photo Scoring was trained on Instagram photos.
Stock Photo Scoring estimates relevance and compliance with stock image trends;
UGC Photo Scoring evaluates the overall attractiveness of images based on their technical characteristics: brightness, contrast, crop, etc.
Stock Photo Scoring estimates properly only stock photos;
UGC Photo Scoring estimates correctly both amateur and professional shots taken on the smartphone or SLR cameras.
Stock Photo Scoring can be useful to designers and stock authors;
UGC Photo Scoring can be useful to developers of search engines and mobile apps.
