Smartphone Bi-Camera System for Backwards Photometric Correction of a 3D Gaussian Splatted Render

Abstract. Multi-camera systems are essential for multi-view stereo, enabling accurate depth estimation and 3D reconstruction through parallax. However, traditional bi-camera setups, typically forward-facing, are limited in capturing diverse environmental perspectives, particularly in autonomous driving, where vehicle motion is restricted. To overcome these limitations, we propose a novel no-parallax bi-camera system that combines a forward-facing and a backward-facing camera on a smartphone setup. This configuration simulates 360° spatial coverage and enhances 3D reconstruction by photometrically correcting forward-facing Gaussian-splatted renders using the backward-facing camera in a loosely-coupled manner, rather than relying on multi-view images in a conventional bundle adjustment. Our results show that incorporating a backward-facing image significantly improves the quality of backward renders, effectively reducing artifacts and mitigating overfitting seen in forward-facing-only images, resulting in a more accurate and refined 3D representation of the environment.