We propose to use a computer vision method to surface aesthetically pleasant pictures from the immense pool of near-zero-popularity items.
[@ICWSM 2015, Oxford]
Have a look at these two Flickr pictures. What’s the difference between them? They are both aesthetically pleasant photos, however their social exposure is very different. Picture A received a lot of attention as showed by the high number of favorites, comments, and views while picture B has very low social signals. These are examples of respectively popular and unpopular content.
Considering the 100M creative commons Flickr dataset, we found that only 2% of the pictures have more than 5 favorites while the remaining 98% have five or less. The vast majority of the content lies below the surface of attention. While exploratory interfaces like the Flickr Explorer page tend to promote appealing and popular images, what happens to the unpopular content? Is it possible that all those 98M pictures do not deserve attention?
In our work instead of focusing on the aspects that make an item popular we flip the perspective and we take a look at the quality of the unpopular content trying to surface the hidden gems of Flickr pictures. To infer image aesthetic quality we resort to crowdsourcing.
We sampled 10K images within 4 topical categories (nature, urban, people and animals). Because we want to see how quality varies with popularity within each category we randomly sample popular, average popular and unpopular pictures. We design a crowdsourcing task where a user has to evaluate how beautiful is an image on a 5-grade scale. We ask at least 5 independent judgments per image.
Even if there is a positive correlation between popularity and beauty (that means on average popular pictures are perceived as appealing) there are popular photos that have low aesthetic value. This is not surprising since popularity is not driven only by intrinsic quality as shown in the past. More importantly, several pictures with near-zero favorites have been judged as very appealing. Their relative amount is very low, making any random sampling strategy totally ineffective.
Social signals like comments or tags are most of the time ineffective since unpopular items rarely receive social feedback that means you cannot use tags or comments to surface them. A possible solution is to look at the pixels. To surface the high quality content we use the crowdsourced dataset to learn a computational aesthetic framework based on compositional visual features. Such framework is able to automatically score images according to their aesthetic value.
With this algorithm we ranked 9M unpopular Flickr photos. We then went back to the crowd and we asked the workers to judge the top 200 rated photos, the most beautiful according to our algorithm. Consistently across categories, the perceived beauty of the surfaced images is comparable to the most popular photos.