Computer Vision News - September 2020

3 Dario Pavllo 51 generate a car, you have to draw the outline or shape of it,” he explains. “That’s fine for simple objects but if you want to draw animals or more complicated objects it presents a limitation. To improve on that, we need to add a step of 3D reasoning because the 3D images we see are actually 3D projections of 3D scenes. Currently, I’m working on integrating this 3D reasoning into generative models so that would be the next step for this work.” He adds that he has seen plenty of examples where models have failed or produced unexpectedly funny results, as you can see in the images of giraffe, zebra, sheep and dog… In terms of computer vision , the model uses GANs , which are the state of the art for generative models and achieve sharp images. The work uses an architecture called SPADE , which was proposed at CVPR 2019 and is also the state of the art, although other architectures could be used with this approach. “We need to add a step of 3D reasoning because the 3D images we see are actually 3D projections of 3D scenes.” Best of ECCV 2020