Panoramic and Omnidirectional Vision Papers

Proposes a learned gradient descent approach for producing multiplane images (MPIs) from sparse input views, enabling accurate scene reconstruction including occlusions and reflections.

Full pipeline to capture and render immersive light field video using a 46-camera hemispherical rig for volumetric rendering.

Portable capture and rendering system for panoramic light field stills with disk-based blending and VP9 compression for VR.

Details the camera rigs and software pipeline behind Google’s immersive light field VR experience for SteamVR.

PanFusion combines Stable Diffusion with panoramic-aware attention for generating 360° images from text prompts.

360DVD presents a diffusion model with a 360-adapter for generating controllable panoramic video from text or prompts.

An encoder-decoder architecture with distortion-aware transformer layers designed for omnidirectional tasks.

Framework to generate full 360° video from a standard camera anchor using dual-branch denoising and antipodal masking.

PanoDiff pipeline to stitch a full panorama from sparse, unregistered narrow-FOV images using angle prediction and diffusion.

Proposes circular convolution and transposed convolution layers for panoramic imagery with wrap-around padding.

Applies CNN-based super-resolution to 360° images, enhancing quality especially around equatorial regions using weighted loss.

Proposes spatially adaptive filters that account for equirectangular distortion in omnidirectional image processing.

Unified geometric model using conformal algebra to treat catadioptric and panoramic imaging systems in robotics.

Synthesizes novel 3D views from a single indoor 360° panorama using mesh reconstruction and Gaussian splatting.

Lightweight network that completes sparse depth input into dense maps using panoramic RGB and dynamic feature fusion.

Tutorial covering state-of-the-art 3D reconstruction methods using panoramic imagery in indoor scenes.

Summarizes 3D modeling methods from the HPCCN Spoke-9 project including floorplans, stereo, and structured geometry pipelines.

Presents UBotNet for depth and normal prediction on real-world omnidirectional data using synthetic-to-real transfer.

Combines structure-from-motion and deep classification to document city centers using low-cost 360° video capture.

Benchmark dataset and tools for cross-device 360° localization with equirectangular and pinhole image queries.

Optimizes spherical SDFs structured in adaptive binoctrees to reconstruct scenes from short 360° video sweeps.

Models spherical signals using harmonic positional encoding for compact implicit neural representation of panoramic data.

Constructs panoramic NeRFs from a single 360° image using depth prediction and RGB-D inpainting to enable roaming.

Splatting method tailored for wide-baseline 360° data using spherical cost volumes and bi-projection encoders.

Leverages indoor layout priors to initialize 3D Gaussians more effectively in equirectangular space for VR roaming.

Jointly optimizes camera parameters and splatting for uncalibrated omnidirectional inputs without prior intrinsics.

Includes lens distortion modeling and cubemap resampling to reconstruct accurate views from fisheye imagery.

Sparse-view 360° reconstruction using uncertainty-aware optimization and stereo bootstrapping in layered splatting.

Describes a hemispherical camera rig using Yi 4K cameras to produce immersive 6-DoF video content.

Facebook’s design for a synchronized multi-view 6DoF camera with software integration for immersive content.

An advanced plenoptic system integrating RGB, IR, and depth sensors for accurate 3D capture in robotics and vision.

Stereo depth estimation using a recurrent matching module based on RAFT, achieving high accuracy on Scene Flow and KITTI.

Augments stereo panoramas with depth to enable parallax and VR navigation in 6-DoF from a single panoramic pair.

Stereo depth estimation for omnidirectional image pairs using recurrent refinement and spherical cost volumes.

Learns stereo depth on omnidirectional images via concentric spherical warping and 3D convolutions over the sweep volume.

A spherical stereo depth model using learnable disparity cost volumes tailored for top-bottom 360° image pairs.

Uses spherical convolution and dual-representation teacher-student training to estimate depth from a single panorama.

Synthesizes stereoscopic environments from a single 360° image, enabling immersive navigation using MCOP images.

Benchmarks GPT-4o and others on spatial reasoning tasks in omnidirectional scenes using OSR-Bench dataset.

Comprehensive survey of deep learning in omnidirectional vision covering datasets, methods, and future challenges.