Developers from Belgium and the USA presented an algorithm that transforms selfies into portraits made by a third party photographer. To do this, the system analyzes human posture in the picture with a third-party algorithm selects the most similar portrait pose and restore the image already with her. A Preprint of the article published on arXiv.org.
Front camera modern smartphones now allow you to do pretty high quality selfie without third-party assistance and tripods. It is still another problem — the broken song: selfie look less natural than the usual portrait shots (they, for example, is often seen outstretched arm), as well as using them rarely manages to capture himself, or as well as the background.
At least one problem — an unnatural posture for the selfie can be corrected using additional processing: for example, crop so that you could not see her outstretched arm. You can do the same with the help of computer vision algorithms and one such algorithm was presented by the researchers under the guidance of the Lijiang MA (Ma Liqian) from the Catholic University of Louvain in Belgium.
At the heart of their system — algorithm DensePosethat automatically determines the pose of the subject (last year, for instance, used to create interactive computer games with live players). Two-dimensional projection of the pose is then compared to the rest of the poses in the database: using the method of k-nearest neighbors algorithm finds natural posture that is most similar to the original at the position of the upper part of the torso, it draws and inserts the frame back.