3D Outdoor Scene Reconstruction from a Single Equirectangular Panoramic Image

Abstract

3D Outdoor scene reconstruction aims to recreate the outdoor scene using multiple images or videos and the camera parameters for mixed reality simulation and novel scene generation. However, the acquisition of various images or videos can prove to be costly and also needs a Structure-from-Motion pipeline for the pre-processing which happens to be quite slow due to the need for an accurate representation of matched features. We propose a novel approach for 3D reconstruction that utilizes a single equirectangular panoramic image which encompasses a 360-degree view for addressing data acquisition issues and then leverages a dense stereo model(DUSt3R) for recovering camera parameters. In addition, to maximize the visual information recovered from the panorama image, we propose a novel interpolation technique for the conversion of the panorama image into a cube map. Both qualitative and quantitative results show that our method can generate better 3D reconstruction for outdoor scenes. We also perform an analysis on using multiple panoramas with translation from Google Street View for 3D outdoor scene reconstruction.