Skip to main navigation Skip to search Skip to main content

End-to-end framework for panoramic indoor style transfer: From photorealistic generation to immersive WebXR exploration

  • Muhammad Tukur*
  • , Sara Jashari
  • , Uzair Shah
  • , Mahmood Alzubaidi
  • , Fabio Bettio
  • , Giovanni Pintore
  • , Enrico Gobbetti
  • , Jens Schneider
  • , Marco Agus
  • *Corresponding author for this work
  • Hamad bin Khalifa University
  • Visual and Data-Intensive Computing
  • CRS4

Research output: Contribution to journal › Article › peer-review

Abstract

We propose PanoStyleSuite, a unified end-to-end framework for panoramic indoor style transfer that integrates photorealistic generation, real-world adaptability, and immersive WebXR-based exploration. Starting from a single 360° panorama, our pipeline employs a multitask dense prediction transformer to infer semantic segmentation, depth, albedo, shading, and layout—providing geometry- and photometry-aware cues for shading-independent stylization. A geometry-aware GAN then applies photorealistic style transfer constrained by depth-, layout-, and edge-consistency losses, while an enhanced intrinsic decomposition model explicitly handles both obscurance and highlight components to ensure illumination-stable stylization. Super-resolution further restores high-frequency details suitable for VR deployment. Building upon this foundation, we introduce an immersive WebXR extension that converts stylized panoramas into stereoscopic multi-center-of-projection views, enabling head-tracked VR exploration. A semantic-aware style embedding, computed from per-class albedo-driven descriptors, supports metric-based interactive in-VR recommendation and real-time style selection. Experiments on Structured3D dataset demonstrate state-of-the-art performance in perceptual fidelity (PSNR, SSIM) and stylistic realism (ArtFID), while qualitative evaluations on Stanford2D3D dataset confirm robust in-the-wild generalization enabled by multitask inference. A VR-based user study with 21 participants indicates that participants assigned higher ratings to the proposed geometry-aware variants across five perceptual dimensions; one-way ANOVA reveals statistically significant differences for Visual Quality, Semantic Awareness, Style Appeal, Seamlessness, and Overall Satisfaction (all (Formula presented) ), underscoring the critical importance of geometry-aware constraints for perceptual realism in immersive environments. Overall, PanoStyleSuite provides the first fully integrated solution spanning photorealistic stylization, real-world adaptability, and immersive WebXR interaction—advancing applications in virtual staging, interior design, real-estate visualization, and metaverse-oriented environments.

Original languageEnglish
Article number106182
JournalImage and Vision Computing
Volume175
DOIs
Publication statusPublished - Nov 2026

Keywords

  • 3D virtual staging
  • Dense prediction
  • Generative adversarial network (GAN)
  • Immersive recommendation system
  • Panoramic indoor scenes
  • Style similarity ranking
  • Style transfer

Fingerprint

Dive into the research topics of 'End-to-end framework for panoramic indoor style transfer: From photorealistic generation to immersive WebXR exploration'. Together they form a unique fingerprint.

Cite this