A while ago, Google released this heat map image showing the places in the world that are photographed the most, based on the geotagging metadata.
Tourist places in Europe get the most traffic, especially art capitals like Paris and Barcelona, coastal regions of Spain, and picturesque regions in the Alps.
Google's artificial intelligence systems no longer need geotagging to know where a photo was taken. After analyzing huge databases of more than 93 million images, they can now recognize where any photo was taken.
Hmm, it could be that, or it could be that the training dataset disproportionately favored those areas based on the limited amount of data... The article says the training set came from flickr, so it likely heavily favors English-speaking users, or at least westerners (Do they use flickr in India or china?). If there are very few examples from central Africa in the training set, but then it reviews a photograph from that area, what is the likelihood that it accurately tags it with the correct location? Also, it's far more likely that the humans who selected the photos they want to put on the web favored their vacation photos, and not say, a field in Nebraska, even though they have far more pictures of that area... Who is uploading the photos to the portion of the Internet that google is reviewing? What about the countless ones it can't identify?
ReplyDeleteMy guess is that there are far more photos taken in India and China, but because those photos aren't in the dataset, and maybe remain locked in peoples' phones, or placed on alibaba rather than Flickr, this tool vastly over-represents a particular area of the world.
People like to photograph water : oceans, seas, rivers. That is true when I analyse France's map. The yellow lights follow the coastlines and the rivers.
ReplyDelete