Answers are grounded in this paper and cite their sources.
Reconstructing animatable humans from monocular video becomes particularly challenging when avatars are driven by poses beyond those observed during training, as locally generated Gaussian attributes may become inconsistent with articulated body support. To address this challenge, we present PoseRoute, an anchor-structured Gaussian avatar framework with selective pose conditioning. Within articulated anchor–child groups, body-aligned anchors follow skeletal motion, while shared decoders generate local child Gaussians. Pose information is routed only to dynamic child offsets through a root-canonicalized local pose representation, while opacity, appearance, and canonical Gaussian shape are decoded from anchor features without direct pose conditioning. Anchor-relative offset and scale constraints further bound local Gaussian geometry during novel-pose driving. Experiments on People-Snapshot, UBC-Fashion, and ZJU-MoCap demonstrate competitive reconstruction quality with real-time rendering at up to 119 FPS. In paired evaluations against our reproduced GART baseline, PoseRoute improves articulated-boundary fidelity. Controlled root-yaw tests further yield 4.6×, 4.7×, and 10.5× reductions in alpha inconsistency, boundary Chamfer distance, and Gaussian drift, respectively. These results show that structured support with controlled pose-dependent Gaussian generation improves avatar consistency under novel-pose driving.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: