“This is certainly the coolest thing I’ve ever worked on, and it might be one of the coolest things I’ve ever seen.”
My Google Research colleague Jon Barron routinely makes amazing stuff, so when he gets a little breathless about a project, you know it’s something special. I’ll pass the mic to him to explain their new work around capturing multiple photos, then synthesizing a 3D model:
I’ve been collaborating with Berkeley for the last few months and we seem to have cracked neural rendering. You just train a boring (non-convolutional) neural network with five inputs (xyz position and viewing angle) and four outputs (RGB+alpha), combine it with the fundamentals of volume rendering, and get an absurdly simple algorithm that beats the state of the art in neural rendering / view synthesis by *miles*.
You can change the camera angle, change the lighting, insert objects, extract depth maps — pretty much anything you would do with a CGI model, and the renderings are basically photorealistic. It’s so simple that you can implement the entire algorithm in a few dozen lines of TensorFlow.
Check it out in action: