Robust 3D Human Reconstruction System From Sparse Views in Outdoor Environments

Abstract

We propose a novel 3D human reconstruction system designed for diverse settings including outdoor environments. Existing methods face the challenges of portability, robustness, and adaptability to environmental variations, as they often rely on sensor data or handcrafted features for marker detection in camera calibration. Consequently, these methods are often limited in outdoor environments. To overcome these limitations, we employ a simple ball-wand as a marker and leverage visual foundation models for marker detection. Our approach enables precise multi-camera calibration without requiring bulky, sensor-specific calibration tools and dependence on controlled environments. Additionally, we introduce a framework that reconstructs 3D human model from sparse input views using occupancy and neural radiance fields. By using sparse views as input, our method efficiently captures both the geometry and appearance of humans in outdoor scenarios, eliminating the need for numerous cameras. We validated our system through extensive experiments on both synthetic human scan datasets and real-world outdoor datasets. The results demonstrate that our system accurately optimizes camera parameters and reconstructs high-quality 3D human models in both outdoor environments and synthetic scenarios.