“PoseVocab: Learning Joint-structured Pose Embeddings for Human Avatar Modeling” by Li, Zheng, Liu, Zhou and Liu
Conference:
Type(s):
Title:
- PoseVocab: Learning Joint-structured Pose Embeddings for Human Avatar Modeling
Session/Category Title: Motion Recipes and Simulation
Presenter(s)/Author(s):
Moderator(s):
Abstract:
We present a new pose encoding method, PoseVocab, for human avatar modeling. Previous methods usually directly map driving poses to dynamic human appearances through a NeRF MLP, yielding blurry avatars. In contrast, PoseVocab constructs pairs of key poses and learnable pose embeddings to encode high-fidelity human appearances under various poses.
