Examination of Minimal Camera Setups for Volumetric Video Capturing from a Time Perspective

Abstract

This study examines the impact of camera count on training time and quality in volumetric video capture using 3D Gaussian Splatting. Two datasets were processed with varying camera numbers, testing two reduction strategies: removing cameras from the center or edges. Results show that 14 cameras offer an optimal balance, reducing training time by up to 32 hours per minute of video (30 fps) with minimal quality loss, while fewer cameras cause significant artifacts. Edge removal proved more efficient, minimizing computational errors.