Leveraging 360° Camera in 3D Reconstruction: A Vision-Based Approach

Abstract

In this paper, we present a novel vision-based approach for 3D reconstruction using a single 360° camera, aiming to offer a simplified and accessible solution for various consumer-oriented applications. Consumer-grade 360° cameras have gained significant popularity due to their affordability and ease of use. However, traditional methods for 3D reconstruction often require complex setups with multiple cameras or expensive hardware such as Light Detection and Ranging (LiDAR). Our approach addresses the challenges associated with 360° cameras by converting the distorted Equirectangular Projection (ERP) into four perspective views, allowing compatibility with deep learning models trained on undistorted perspective images. We leverage Visual Simultaneous Localization and Mapping (VSLAM) techniques for camera pose estimation and employ a standard 3D reconstruction pipeline for generating detailed 3D mesh representations of the indoor environment. Through experimental evaluation, we compare the performance of 360° cameras with traditional perspective cameras in 3D reconstruction and analyze the accuracy and performance of our vision-based approach. Our findings demonstrate the potential of using 360° cameras for constructing high-quality models and facilitating efficient data collection for 3D reconstructions, opening up new possibilities for various consumer-oriented applications in multiple fields.