Computer Vision News - September 2020

35 Poster Presentation Zak Murez (Magic Leap) Atlas: End-to-End 3D Scene Reconstruction from Posed Images Author Zachary Murez told us that he is excited about the possibility of combining Atlas with ideas from another ECCV2020 paper that we reviewed on Monday: “NeRF Representing Scenes as Neural Radiance Fields for View Synthesis” (read on page 32), allowing for increased resolution, reduced memory, and self-supervised training. “The network architecture bears a lot of resemblance to prior cost volume based stereo depth networks , ” Zachary added “but the key differences are that features are backprojected into a canonical voxel volume instead of planes of a view frustum, and voxel occupancy is regressed instead of depth probability.” Best of ECCV 2020