This content was created by LARRUZ. This content was uploaded by LARRUZ. This is a YouTube video titled "Algorithmic Beautification of Selfies | Two Minute Papers #159" (published June 5, 2017). It is a short explanatory monologue by the Two Minute Papers host describing a computer-vision method for reconstructing a 3D face and estimating camera distance from a single selfie to correct perspective distortion (e.g., nose and forehead magnification). The speaker explains the fitting process using facial landmarks/hotspots (with additional landmarks added for better quality), notes the method runs quickly (under ~5 seconds), and describes applications: changing perceived camera distance, rotating the 3D face, producing difference images for error localization, stereoscopy, and an interactive online demo. The host also suggests combining this with a neural-network quality scorer to automatically optimize the “perfect selfie,” and closes with a viewer call-to-action (try demo, comment, subscribe). Metadata: YOUTUBE content, Two Minute Papers series, created 2017-06-05 (last edited 2026-01-02).
Chapters
Dear fellow scholars, this is two minute papers with Today we are going to talk about a rigorous scientific topic, none other than the creation of the perfect selfie photo. By definition, selfies are made by us, which means that these are typically short range photos. And due to the perspective distortion of the camera lens, we often experience unpleasant effects like the heavy magnification of the nose and the forehead. And get this. This technique enables us to take a photo and after that, edit the perceived camera distance for it without changing anything else. Basically algorithmic beautification. This technique works the following. We analyze the photo and try to figure out how distant the camera was when the photo was taken. Then we create a digital model of the perspective camera and create a 3D model of the face. This is a process that mathematicians like to call fitting. It means that if we know the optics of perspective cameras, we can work backwards from the input photo that we have and find an appropriate setup that would result in this photo. Then we will be able to adjust this distance to even out the camera lens distortions. But that's not all, because as we have the digital 3D model of the face, we can do even more. For instance, we can also rotate it around in multiple directions. To build such a 3D model, we typically try to locate several well recognizable hotspots on the face such as the chin, eyebrows, nose stem, the region under the nose, eyes and lips. However, as these hotspots lead to a poor 3D representation of the human face, the authors added a few more of these hotspots to the detection process. This still takes less than five seconds. Earlier, we also talked about a neural network based technique that judged our selfie photos by assigning a score to them. I would absolutely love to see how that work would react to a before and after photo that comes from this technique. This way we can formulate this score as a maximization problem and as a result, we could have an automated technique that truly creates the perfect selfie photo through these warping operations. The best kind of evaluation is when we let reality be our judge and use images that were taken closer or farther away and compare the output of this technique against them. These true images bear the ground truth label throughout this video. The differences are often barely perceptible and to provide a better localization of the error, some difference images are shown in the paper. If you are into stereoscopy, there is also an entire section about that as well. The authors also uploaded an interactive version of their work online that anyone can try free of charge. So as always, your scholarly before and after selfie experiments are more than welcome in the comment section. Whether you are already subscribed to the series or just subscribing now, which you should absolutely do, make sure to click the bell icon to never miss an episode. We have lots of amazing works coming up in the next few videos. Hope to see you there again. Thanks for watching and for your generous support and I'll see you next time. Ra.
