# PaGeR Framework Brings State-of-the-Art 3D Geometry Estimation to Single Panoramic Images

_Research · published 2026-06-04_

Researchers from ETH Zurich, Google, Meta, and Athlence Sports have introduced PaGeR (Panoramic Geometry Reconstruction), a framework that adapts perspective-image 3D foundation models to the panoramic domain. Rather than redesigning the architecture, PaGeR projects equirectangular panoramas into six-face cubemaps and feeds them to a pre-trained multi-view transformer — specifically Depth Anything 3 with a ViT-Giant backbone — treating the cube faces as a multi-view input set. A single forward pass produces scale-invariant depth, metric depth, surface normals, and sky masks from both perspective and 360-degree images, achieving state-of-the-art results on indoor and outdoor benchmarks.

To support training and evaluation, the team releases two new datasets: PanoInfinigen, a synthetic collection of 77,000 panoramas rendered at 4K with pixel-perfect ground-truth depth and normals, and ZüriPano, a real-world outdoor benchmark of 100 LiDAR-scanned urban panoramas captured across Zürich at 8K resolution with an effective depth range of 130 metres. The authors report that existing perspective foundation models degrade significantly on large-scale outdoor scenes, while PaGeR maintains strong zero-shot performance across both indoor and outdoor environments.

## Sources
- [pager360.github.io](https://pager360.github.io/)
---
Canonical: https://genbuzz.news/posts/pager-framework-brings-state-of-the-art-3d-geometry-estimation-to-single-panoramic-images