Abstract
We introduce 3D Moments, a new computational photography effect. As input wetake a pair of near-duplicate photos, i.e., photos of moving subjects fromsimilar viewpoints, common in people's photo collections. As output, we producea video that smoothly interpolates the scene motion from the first photo to thesecond, while also producing camera motion with parallax that gives aheightened sense of 3D. To achieve this effect, we represent the scene as apair of feature-based layered depth images augmented with scene flow. Thisrepresentation enables motion interpolation along with independent control ofthe camera viewpoint. Our system produces photorealistic space-time videos withmotion parallax and scene dynamics, while plausibly recovering regions occludedin the original views. We conduct extensive experiments demonstrating superiorperformance over baselines on public datasets and in-the-wild photos. Projectpage: https://3d-moments.github.io/