Abstract
Spherical images taken in all directions allow the full surroundings of a subject to be represented, providing an immersive experience. Generating a spherical image from a single normal-field-of-view (NFOV) image is convenient but challenging due to the high degree of freedom in generating unobserved regions. We propose a method that leverages scene symmetry—rotational, planar, and asymmetric—as a structural prior. By incorporating symmetry intensity as a latent variable in a Conditional Variational Autoencoder (CVAE) and using circular shifts and flips of latent features, we control the generated regions. Our results demonstrate that this approach produces plausible, seamless spherical images with reduced reconstruction errors.
Introduction
Spherical (360°) images are essential for virtual reality and immersive content, yet capturing them typically requires specialized hardware or complex stitching. Generating a full 360° environment from a single standard photograph (NFOV) would significantly lower the barrier for VR content creation. However, this is an extreme image completion task. Standard models often fail to maintain spherical continuity, leading to visible seams, and struggle to hallucinate the vast unobserved areas without a strong structural prior.
Methodology: Controlling Symmetry in Latent Space
The proposed method utilizes a Conditional Variational Autoencoder (CVAE) architecture designed for equirectangular projections. The core innovation lies in treating scene symmetry as a controllable latent variable.
Symmetry as a Structural Prior
Most real-world scenes, such as architectural interiors or courtyards, exhibit global symmetry. The model introduces a symmetry intensity vector $s$ to control five types of symmetry: 90°, 180°, and 270° rotational symmetry, and 0° and 90° plane symmetry.
Geometric Transformations and Circular Padding
To enforce these symmetries, the model applies explicit geometric operations—horizontal circular shifts for rotational symmetry and flips for plane symmetry—directly to the CNN feature maps. Furthermore, to ensure the generated 360° image is perfectly continuous at the boundaries, standard zero-padding is replaced with circular padding in all convolutional layers. This architectural tweak eliminates the edge discontinuities common in other equirectangular generation methods.
Experimental Results
The model was trained on the SUN360 dataset, using 50,000 images for training and 10,000 for testing. Input NFOV images were synthetically generated with fields-of-view ranging from 30° to 120°.
Quantitative Performance
The proposed method significantly outperformed state-of-the-art baselines like PICNet and 360IC. In generation tasks, it achieved a Frechet Inception Distance (FID) of 24.4 (vs. PICNet's 28.3). In reconstruction, it reached a PSNR of 14.3, compared to 12.5 for PICNet.
Qualitative Control and Seamlessness
Beyond metrics, the model allows users to manually adjust the symmetry latent variable during inference, enabling smooth transitions from asymmetric to highly symmetric outputs. Qualitative evaluations confirmed that the circular padding successfully produced seamless equirectangular projections, a key advantage over previous methods.
Conclusion
By explicitly modeling scene symmetry and adapting CNN architectures for spherical geometry, this work provides a robust framework for generating immersive 360° environments from limited visual data. While challenges remain in controlling local object placement and generalizing to highly unusual scenes, the use of symmetry as a global structural prior marks a significant step forward in unconstrained image completion.
Summary updated