A Fast End-to-End Method with Style Transfer for Room Layout Estimation

摘要

In this paper, we focus on the problem of estimating the spatial layout of indoor scenes from a single RGB image. Earlier solutions to this task largely rely on hand-crafted features and vanishing point detection, and recent practices usually utilize a fully convolution network (FCN) to achieve better results. However, existing works suffer from time consumption and sufficient usage of semantic information in each pixel. To tackle these issues, based on the FCN architecture, we propose a fast end-to-end method for room layout estimation. For the first time, style transfer is introduced for data enhancement, which not only extends the amount of training dataset, but also makes the pixels in images more discriminative, especially in highly cluttered rooms. Besides, we further improve the accuracy by fusing the output from multiple networks. Extensive evaluations on the public large-scale scene understanding challenge (LSUN) dataset demonstrate that our proposed method builds a new state-of-the-art result in terms of speed and accuracy.