Algorithm detects relationships on social media; finds users in untagged photos
TORONTO – A simple keyword search on Facebook can reveal a lot of information about a user; but what if there were a tool intelligent enough to not only determine your relationship with other users but reveal photos of you online that you didn’t know existed?
A new algorithm developed by University of Toronto engineering professor Parham Aarabi may soon allow users to do just that, by using the names and locations of existing photo tags to create a relationship graph between users without using facial-recognition software.
For example, if a father and son are tagged in images over and over again, the algorithm can detect the relationship between the two in photos that are not tagged by evaluating how close the two subjects are standing in the image and how many times they have been tagged together.
Similarly, if a photo of the mother, the son, and the father is tagged – but the father is left untagged – the tool is still able to recognize him based on all of the other tagged images. If the mother was untagged, it would search through other tagged images of her to detect she is likely related.
“The best way to understand an image is to create a mathematical novel of what is in the image and how the objects in the image interact. For example, if you are standing beside your parents and siblings, that connection tells the computer that this is a family photo,” Aarabi told Global News.
“The core idea from this work is that based on the tags that co-appear in images – images that you and your dad are tagged in, for example – we can say that perhaps you and your dad have a close affinity.”
Facebook, for example, uses facial recognition software to help users tag friends or family members in their photos. When a user uploads an image Facebook can suggest tags for the photo based on images you have previously uploaded and the site’s database of images from other users.
But Aarabi believes the tool – dubbed “relational social image search” – will lead to better search capabilities on social networking sites like Facebook or Flickr. This is because the algorithm is based on the number of tags, not the number of photos on a site, making it a more efficient search tool.
“By understanding the people that are near to you we can actually find photos that you were in but not tagged in, or photos that might be relevant to you,” he said.
“If you search for me on Facebook right now you might find about 20 images, but I am actually in about 100 images and with this algorithm all of those images could be found.”
Aarabi, and his former Master’s student Ron Appel who helped develop the tool, plan to start discussions with social networks to see how the algorithm could be used on a site wide basis after the algorithm is presented at the IEEE International Symposium on Multimedia on Tuesday.
“I envision the interface would be exactly like you use Facebook search—for users, nothing would change. They would just get better results,” says Aarabi.
© Shaw Media, 2013