PointHuman: Learning high-fidelity and generalizable human neural radiance fields using guidance of fine-grained semantics-enriched geometry

Jintai Wang, Pengfei Liu, Huidong Feng, Qifeng Dai, Yinglin Zheng, Ming Zeng

Abstract: Highlights•Fine-grained geometry is reconstructed from sparse human views.•A semantic-aware deformation field maps pixels to accurate 3D points.•Spatial encoding from neural points boosts rendering fidelity.•Outperforms baselines on generalizable human NeRF(neural radiance field) tasks.•Supports high-quality rendering under sparse input and unseen poses.

External IDs:dblp:journals/cg/WangLFDZZ25