Rao Shivansh ; Kumar Vikas ; Kifer Daniel ; Giles Lee ; Mali Ankur

Abstract

Given a single RGB panorama, the goal of 3D layout reconstruction is toestimate the room layout by predicting the corners, floor boundary, and ceilingboundary. A common approach has been to use standard convolutional networks topredict the corners and boundaries, followed by post-processing to generate the3D layout. However, the space-varying distortions in panoramic images are notcompatible with the translational equivariance property of standardconvolutions, thus degrading performance. Instead, we propose to use sphericalconvolutions. The resulting network, which we call OmniLayout performsconvolutions directly on the sphere surface, sampling according to inverseequirectangular projection and hence invariant to equirectangular distortions.Using a new evaluation metric, we show that our network reduces the error inthe heavily distorted regions (near the poles) by approx 25 % when compared tostandard convolutional networks. Experimental results show that OmniLayoutoutperforms the state-of-the-art by approx 4% on two different benchmarkdatasets (PanoContext and Stanford 2D-3D). Code is available athttps://github.com/rshivansh/OmniLayout.