ECCV 2020 Daily - Monday

discrete sample grid, but we’re able to represent the same complex scene with only five megabytes . The trade-off is that now you have to do the computation of querying the network at every point in the grid instead of just indexing.” Thinking specifically about the computer vision techniques used in this work, Pratul tells us it uses a small, fully connected deep network to represent a volume. It then uses techniques from traditional graphics like continuous volume rendering to render the network’s represented volume into a new viewpoint. These volume- rendering techniques from graphics are differentiable, so can train all theweights of the network that represent the volume from just the images observed of the scene. “One of the reasons that this works so well is because of this idea of multi- view consistency,” Pratul explains. “If I have a bunch of images and they’re all viewing some point in the world, what is the volume density and the color in any direction at this point? The network will do the best job at reproducing the images if it allocates a lot of volume density to that point. Naturally, the network is encouraged to represent the real world because that’s going tomake the pictures multi-view consistent so that any scene point will show up in all the pictures.” 2 Oral Presentation 6 DAILY M o n d a y “One of the reasons that this works so well is because of this idea of multi-view consistency!”