PaGeR: Panoramic Geometry Reconstruction via Multi-View Foundation Models

Researchers have rolled out a new system called PaGeR, which stands for Panoramic Geometry Reconstruction. This framework is designed to adapt pre-trained 3D models—meant for regular perspective images—so they can work with 360-degree panoramic images. What’s cool is that it can effectively predict things like depth and surface normals from both types of images in one go. By training with both perspectives and panoramic images while only tweaking the model a bit, PaGeR achieves impressive accuracy in geometry estimation. This breakthrough significantly boosts the capability of traditional 3D reconstruction models, allowing them to fully reconstruct scenes from a single panoramic image using a slightly modified pre-trained transformer.