UnderPressure: Deep Learning for Foot Contact Detection, Ground Reaction Force Estimation and Footskate Cleanup

Human motion synthesis and editing are essential to many applications like video games, virtual reality, film post-production. However, they often introduce artefacts in capture data, which can be detrimental the perceived realism. In particular, footskating is a frequent disturbing artefact, requires knowledge of foot contacts cleaned up. Current approaches obtain contact labels rely either on unreliable threshold-based heuristics or tedious manual annotation. this article, we address automatic label detection from data with deep learning based method. To end, first publicly release Under Pressure, novel database labelled pressure insoles serving as reliable ground. Then, design train neural network estimate ground reaction forces exerted feet then derive accurate labels. The evaluation our model shows that significantly outperform heuristic height velocity thresholds approach much more robust when applied sequences suffering perturbations noise footskate. We further propose fully workflow for footskate cleanup: derived estimated forces. removed by solving constraints through an optimisation-based inverse kinematics (IK) ensures consistency Beyond cleanup, both method could help improve forces, including dynamics problems reconstruction models character animation. Our implementation, pre-trained well links found at github.com/InterDigitalInc/UnderPressure.