Reproduce “HAFMat: Hybrid Priors Guided Adaptive Fusion for Single-Image Human Material Estimation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yu Jiang, Jiahao Xia, Jiongming Qin, Jianchi Sun, Chunxia Xiao
HAFMat presents a framework for estimating five Physically Based Rendering (PBR) material maps from a single human image by leveraging hybrid guidance maps and an adaptive multi-layer fusion mechanism. This approach achieved the highest overall PSNR for both material estimation and relighting metrics on the OpenHumanBRDF dataset, yielding more accurate material properties and realistic rendered appearances.
Key insights
- Diverse guidance cues for material estimation possess heterogeneous characteristics, with some providing local texture details and others offering global semantic context.
- An adaptive, multi-layer feature fusion mechanism is crucial for optimally integrating these heterogeneous cues, allowing the model to prioritize different types of information at varying representational depths.
- Explicitly incorporating hybrid prior guidance, including SMPL-derived geometric/structural maps and pre-trained material predictions, significantly mitigates the ill-posed nature of single-image PBR material estimation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.