PIFu for the Real World: A Self-supervised Framework for Reconstructing Dressed Humans from Single-view Images
Kernekoncepter
SelfPIFu proposes a self-supervised framework using depth-guided learning to improve 3D human reconstruction accuracy and robustness.
Resumé
SelfPIFu introduces a novel self-supervised network named SelfPIFu that leverages in-the-wild images for improved reconstructions. The framework focuses on utilizing depth information to enhance the accuracy of 3D human shape predictions. By incorporating volume-aware and surface-aware signed distance fields (SDF) learning, SelfPIFu enables self-supervised learning without the need for ground truth mesh data. This approach results in superior reconstructions compared to existing methods like PIFuHD and ECON, especially on real-world images with diverse poses and garment complexities. The proposed method demonstrates significant advancements in reconstructing detailed geometric shapes of clothed humans from single-view images.
"Extensive experiments demonstrate the effectiveness of our self-supervised framework and the superiority of using depth as input."
"Our method excels at reconstructing geometric details that are both rich and highly representative of the actual human."
"Our SDF-based PIFu effectively learns convincing surface details especially for in-the-wild images."