A short YouTube explainer (Two Minute Papers #245) summarizing a paper that uses a convolutional neural network to perform real-time facial alignment and 3D face reconstruction from single 2D images. The algorithm produces high-quality geometry in under 10 milliseconds per image, is robust to lighting, pose, and occlusions (e.g., glasses, hair), and outperforms competitors in both speed (5–8× faster) and quality. The video suggests near-term applications in avatars, movies, and games.
Chapters
Dear fellow scholars, this is two minute papers with Today we have two extremely hard problems on the menu. One is facial alignment and the other is 3D facial reconstruction. For both problems we have an image as an input and the output should be either a few lines that mark the orientation of the jawline, mouth and eyes. And in the other case, we are looking for a full 3D computer model of the face and all this should happen automatically without any user intervention. This is extremely difficult because this means that we need an algorithm that takes a 2D image and somehow captures 3D information from this 2D projection, much like a human would. This all sounds great and would be super useful in creating 3D avatars for Skype calls or scanning real humans to place them in digital media such as feature movies and games. That would be amazing, but is this really possible? This work uses a convolutional neural network to accomplish this, and it not only provides high quality outputs, but it creates them in less than 10 milliseconds per image, which means that it can process a hundred of them every second. That is great news indeed, because it also means that doing this for video in real time is also a possibility. But not so fast, because if we are talking about video, new requirements arise. For instance, it is important that such a technique is resilient against changes in lighting. This means that if we have different lighting conditions, the output geometry the algorithm gives us shouldn't be wildly different. The same applies to camera and pose as well. This algorithm is resilient against all three, and it has some additional goodies. For instance, it finds the eyes properly through glasses and can deal with cases where the jawline is occluded by the hair and or infer its shape when one side is not visible at all. One of the key ideas is to give additional instruction to the convolutional neural network to focus more of its efforts to reconstruct the center regions of the face because that region contains more discriminative features. The paper also contains a study that details the performance of this algorithm. It reveals that it is not only five to eight times faster than the competition, but also provides higher quality solutions. Since these are likely to be deployed in real world applications very soon, it is a good time to start brainstorming about possible applications for this. If you have ideas beyond the animation, movies and games line, let me know in the comment section. I will put the best ones in the video description. Thanks for watching and for your generous support and I'll see you next time.
