A Unified Framework for Real Time Motion Completion

Motion completion, as a challenging and fundamental problem, is of great significance in film game applications. For different motion completion application scenarios (in-betweening, in-filling, blending), most previous methods deal with the problems case-by-case methodology designs. In this work, we propose simple but effective method to solve multiple under unified framework achieves new state-of-the-art accuracy on LaFAN1 (+17% better than sota) evaluation settings. Inspired by recent success self-attention-based transformer models, consider sequence-to-sequence prediction problem. Our consists three modules - standard encoder self-attention that learns long-range dependencies input motions, trainable mixture embedding module models temporal information encodes key-frame combinations form, perceptual loss for capturing high-frequency movements. can predict missing frames within single forward propagation real-time get rid post-processing requirement. We also introduce novel large-scale dance movement dataset exploring scaling capability our its effectiveness complex