State-of-the-art in deep learning approaches for automatic single-panorama indoor modeling and exploration

A single surround-view panoramic image provides complete coverage of the environment visible from a single viewpoint and inherently supports dynamic exploration, especially when viewed through a head-mounted display. For these reasons, single or linked 360° panoramas have become a widely adopted modality for indoor scene acquisition and virtual tour creation. Despite their popularity, panoramas present inherent limitations, as they only statically represent the captured scene, do not provide explicit 3D architectural structure and geometry, and exhibit minimal parallax due to their single-viewpoint nature, which limits their application capabilities or requires significant modeling efforts to generate missing data. In this survey, we provide an up-to-date integrative overview of recent techniques designed to overcome these challenges, bringing together complementary perspectives from machine learning, computer vision, and computer graphics. After introducing a characterization of the panoramic input and the target geometric, structural, and visual outputs, we discuss the role of reconstruction priors and motivate the choice of deep learning approaches for leveraging large-scale data to infer hidden information. Next, we outline the main sub-problems involved in lifting a 360° image into a structured, explorable model and review advances in single-view pixel-wise geometric and semantic analysis, single-view indoor layout estimation, localization and multi-room reconstruction from very sparse coverage, novel view synthesis for providing parallax, and immersive model exploration. We then discuss the emergence of both general-purpose and 360°-specific vision foundation models for single-panorama indoor modeling and exploration. Finally, we highlight practical applications and identify open research directions.