Core Concepts
CLOAF eliminates self-intersections in body pose estimation using ODEs, improving accuracy without compromising reconstruction quality.
Abstract
Abstract: CLOAF uses Ordinary Differential Equations to prevent self-intersections in body pose estimation.
Introduction: Current methods struggle with self-intersections in body poses, impacting realism and applicability.
Related Work: Various approaches attempt to address self-intersections but fall short of complete elimination.
Collision-Aware Optimization: Post-processing methods aim to remove self-intersections but are not fully differentiable.
Motion Field Integration: CLOAF integrates ODE-based fields to prevent self-intersections effectively.
Inverse Kinematics: CLOAF combines ODEs with parametric models for accurate pose recovery without self-intersections.
Method: CLOAF integrates motion fields to eliminate self-intersections efficiently and improve body trajectory computations.
Experiments: Training on AMASS dataset, evaluation on 3DPW dataset shows improved accuracy and reduced collision rates.
Stats
HMR2.0 recovers bodies with self-intersections in 39.2% of frames of the 3DPW-test set.
A recent post-processing method brings this down to 9.2%.
CLOAF drops this number all the way to zero.