Efficient Instance Segmentation of Panoramic Images of Indoor Scenes

This paper addresses the issue of efficient 2D instance segmentation of 360$^\circ$ images of indoor scenes. In particular, we study the use of equirectangular convolutions and the impact of different approaches to handle wrap-around areas. We consider the use of Mollweide projection as a representation for performing segmentation, and we provide a toolchain to prepare the Matterport panoramic images for use in workflows designed for COCO-style annotated datasets. The results show no significant differences between using regular and equirectangular convolutions. While the Mollweide projection allows for segmentation of otherwise missed objects, the overall results do not outperform analysis on equirectangular projection.