Publication
Dense Orientation Recovery via Lightweight Graph Attention for Real-Time Avatar Animation from Markerless Tracking
Muhammad Saif Ullah Khan; Chen-Yu Wang; Tim Prokosch; Michael Lorenz; Bertram Taetz; Didier Stricker
In: SCA 2026: ACM SIGGRAPH / Eurographics Symposium on Computer Animation - Posters. ACM SIGGRAPH / Eurographics Symposium on Computer Animation (SCA-2026), July 7-10, Barcelona, Spain, The Eurographics Association, 7/2026.
Abstract
Markerless human tracking systems commonly provide 3D joint positions, whereas real-time avatar animation requires joint
orientations for skeletal skinning. This position-to-orientation problem is underconstrained, especially because twist around
bone axes is not determined by joint locations. We present IK-GAT, a lightweight graph-attention model that recovers full-
body joint orientations from root-centered 3D joints in a single forward pass. Instead of regressing rig-specific local rotations
directly, IK-GAT predicts canonical bone-aligned world rotations defined from the rest-pose skeleton. These predictions are
analytically and exactly converted to standard parent-relative local rotations using the known kinematic tree, making the output
directly usable by animation rigs or SMPL-like body models. The network uses skeletal graph attention, a continuous 6D
rotation representation, and geodesic supervision on SO(3). On AMASS, IK-GAT reduces test MPJAE from 9.06◦ for a dense
Transformer to 7.43◦ while using 8.6× fewer parameters. On an Unreal Engine mannequin benchmark, it reduces local-space
MPJAE from 15.61◦ to 12.83◦ and directly drives production-rig avatars without iterative fitting or runtime retargeting.