HorizonNet: Learning Room Layout with 1D Representation and Pano Stretch Data Augmentation

We present a new approach to the problem of estimating the 3D room layoutfrom a single panoramic image. We represent room layout as three 1D vectorsthat encode, at each image column, the boundary positions of floor-wall andceiling-wall, and the existence of wall-wall boundary. The proposed network,HorizonNet, trained for predicting 1D layout, outperforms previousstate-of-the-art approaches. The designed post-processing procedure forrecovering 3D room layouts from 1D predictions can automatically infer the roomshape with low computation cost - it takes less than 20ms for a panorama imagewhile prior works might need dozens of seconds. We also propose Pano StretchData Augmentation, which can diversify panorama data and be applied to otherpanorama-related learning tasks. Due to the limited data available fornon-cuboid layout, we relabel 65 general layout from the current dataset forfinetuning. Our approach shows good performance on general layouts byqualitative results and cross-validation.