# AI Makes 3D Models From Photos | Two Minute Papers #122 ## Метаданные - **Канал:** Two Minute Papers - **YouTube:** https://www.youtube.com/watch?v=HO1LYJb818Q - **Дата:** 25.01.2017 - **Длительность:** 3:01 - **Просмотры:** 72,166 - **Источник:** https://ekstraktznaniy.ru/video/14722 ## Описание The paper "Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling" and its source code is available here: http://3dgan.csail.mit.edu/ https://arxiv.org/pdf/1610.07584v2.pdf More about generative adversarial networks (and some explanations): Image Editing with Generative Adversarial Networks - https://www.youtube.com/watch?v=pqkpIfu36Os Image Synthesis From Text With Deep Learning - https://www.youtube.com/watch?v=rAbhypxs1qQ WE WOULD LIKE TO THANK OUR GENEROUS PATREON SUPPORTERS WHO MAKE TWO MINUTE PAPERS POSSIBLE: Sunil Kim, Daniel John Benton, Dave Rushton-Smith, Benjamin Kang. https://www.patreon.com/TwoMinutePapers Subscribe if you would like to see more of these! - http://www.youtube.com/subscription_center?add_user=keeroyz Music: Dat Groove by Audionautix is licensed under a Creative Commons Attribution license (https://creativecommons.org/licenses/by/4.0/) Artist: http://audionautix.com/ Thumbnail image background credit: https://pi ## Транскрипт ### [] Dear Fellow Scholars, this is Two Minute Papers with Károly Zsolnai-Fehér. What if we tried to build a generative adversarial network for 3D data? ### Randomly Sampled Shapes [0:06] This means that this network would work not on the usual 2 dimensional images, but instead, on 3 dimensional shapes. So, the generator network generates a bunch of different 3 dimensional shapes, and the ### 3D Generative Adversarial Network [0:23] basic question for the discriminator network would be - are these 3D shapes real or synthetic? The main use case of this technique can be, and now watch closely, taking a photograph from a piece of furniture, and automatically getting a digital 3D model of it. Now it is clear for both of us that this is still a coarse, low resolution model, but it is incredible to see how a machine can get a rudimentary understanding of 3D geometry in the presence of occlusions, lighting, and different camera angles. That's a stunning milestone indeed! It also supports interpolation between two shapes, which means that we consider the presumably ### Interpolation in Latent Space [1:00] empty space between the shapes as a continuum, and imagine new shapes that are closer to either one or the other. We can do this kind of interpolation, for instance between two chair models. But the exciting thing is that no one said it has to be two objects of the same class. So we can go even crazier, and interpolate between a car and a boat. Since the technique works on a low-dimensional representation of these shapes, we can also perform these crazy algebraic operations between them that follow some sort of intuition. ### Arithmetic in Latent Space [1:39] We can add two chairs together or subtract different kinds of tables from each other. Absolute madness. And one of the most remarkable things about the paper is that the learning took place on a very limited amount of data, not more than 25 training examples per class. One class we can imagine as one object type, such as, chairs, tables or cars. The authors made the source code and a pretrained network available on their website, the link is in the video description, make sure to have a look! I am so happy to see breakthroughs like this in machine learning research. One after another in quick succession. This work is surely going to spark a lot of followup papers, and we'll soon find ourselves getting extremely high quality 3D models from photographs. Also, imagine combining this with a 3D printer! You take a photograph of something, run this algorithm on it, and then print a copy of that furniture or appliance for yourself. We are living amazing times indeed! Thanks for watching and for your generous support, and I'll see you next time!