Augmented Reality Enhanced: 3D Crowd Reconstruction from a Single Viewpoint
Xiaohan Sun, Xiao‐Song Yang · 2024
Reconstructing human figures from a single view-point has long intrigued researchers, particularly for augmented reality (AR) applications. While significant progress has been made in single-human body reconstruction, densely populated scenes with substantial occlusions pose complex challenges. This paper introduces 3DCrowd+, an advanced two-stage methodology for 3D reconstruction of human meshes in crowded environments. Building on the 3DCrowdNet framework, our model refines HRNet 2D pose estimation and integrates Lite-HRNet with Shuffle Block and CoordAttention modules, achieving robust feature extraction and lightweight performance. 3DCrowd+ combines an attention mechanism with a model pruning algorithm, demonstrating high accuracy and efficiency on various datasets. This research bridges the gap between complex crowd scenes and detailed 3D reconstruction, offering a promising solution for precise crowd modeling in AR environments.