Texturized Multi-level Implicit Modelling for High- Resolution 3D Human Digitization: The PIFuHD approach
Author(s):
Mihir Harne, Parth Gorde, Shreya Junagade, Roma Thakur
Keywords:
Abstract
Our 3D human shape estimation network stands out for integrating volumetric feature transformation, merging diverse image features into 3D space to precisely recover surface geometry. Complemented by a rich dataset of 7000 real-world human models, our method, empowered by unique architecture, excels in single-image 3D human model estimation. Addressing challenges in estimating human pose and body shape from 3D scans over time, we introduce PIFuHD Pixel- aligned Implicit Function. PIFuHD enables end-to-end deep learning for digitizing detailed clothed humans from a single image, surpassing prior work with high-resolution reconstructions on the Render people dataset. Moreover, our innovative approach recovers fine details, even on occluded parts, by transforming shape regression into an aligned image-to-image translation problem. Using a partial texture map as input, our method estimates detailed normal and vector displacement maps, enhancing clothing representation on a low-resolution smooth body model. In the landscape of 3D human shape estimation, our multi-level architecture, balancing broad context and high resolution, significantly outperforms existing techniques, leveraging 1kresolution input images for enhanced single-image reconstructions.
Article Details
Unique Paper ID: 163951
Publication Volume & Issue: Volume 10, Issue 11
Page(s): 2219 - 2225
Article Preview & Download
Share This Article
Join our RMS
Conference Alert
NCSEM 2024
National Conference on Sustainable Engineering and Management - 2024