Abstract: With the rapid growth of panorama photography applications such as Android Photo Sphere developed by Google, the convenience of creating panoramic images has made it very popular in recording omni-directional information. In this paper we address the problem of room structure reconstruction from an indoor 360° panorama, and a system combining automatic reasoning and semi-interactive refinement is proposed. The system first calibrates the camera and performs perspective sampling to transform the original panorama into multiple perspective projected images, then automatically predicts a raw indoor structure via candidate evaluation using a linear scoring function based on features extracted from the perspective images. Afterwards, users could utilize the semi-interactive refinement functionality of the system, with operations on portable devices, to further adjust the room structure and models a better 3D indoor structure conveniently.