|
Abstract
|
We evaluate a pose-guided system that predicts a kinematic human representation from sparse two-dimensional joints and renders a polyhedral wireframe. The main evidence comes from matched 50,000-update experiments on two synthetic capture sets, C1 and C2. On C1, FULL improves localization over FK-S3D with accurate input joints, but the Teacher-input differences are not resolved by source-bootstrap intervals. C2 compares DIRECT, FK-S3D, and FULL on 756 images, of which 747 admit Teacher inference. Under the same three-dimensional supervision, DIRECT yields lower localization and elbow/knee angle errors than FK-S3D in both tested seeds and input conditions, but larger bone-length errors. FULL uses the same architecture and FK as FK-S3D with additional supervision. Its comparison with DIRECT therefore changes both parameterization and losses, and their Teacher-input intervals include zero. Supervision deletions support a limited localization benefit from reprojection; rotation supervision affects limb-intersection indicators more consistently than mean joint error. Optional 80-step IK can improve projection fitting while worsening synthetic three-dimensional error under inaccurate observations. A historical evaluation of 16 anime images remains exploratory and lacks three-dimensional ground truth. These results identify supervision-dependent benefits and geometric trade-offs rather than a general accuracy advantage from FK.
|