← ArchivePaper2022
DeepPhase: Periodic Autoencoders for Learning Motion Phase Manifolds
Periodic Autoencoder extracts unsupervised spatial-temporal phase features from large unstructured motion datasets, winning SIGGRAPH 2022 Best Paper.
Abstract
Learning the spatial-temporal structure of body movements is a fundamental problem for character motion synthesis. In this work, we propose a novel neural network architecture called the Periodic Autoencoder that can learn periodic features from large unstructured motion datasets in an unsupervised manner. The character movements are decomposed into multiple latent channels that capture the non-linear periodicity of different body segments while progressing forward in time. Our method extracts a multi-dimensional phase space from full-body motion data, which effectively clusters animations and produces a manifold in which computed feature distances provide a better similarity measure than in the original motion space to achieve better temporal and spatial alignment. We demonstrate that the learned periodic embedding can significantly help to improve neural motion synthesis in a number of tasks, including diverse locomotion skills, style-based movements, dance motion synthesis from music, synthesis of dribbling motions in football, and motion query for matching poses within large animation databases.
How to read this
- Category
- Method: neural architecture for unsupervised motion phase manifolds
- Contributions
- A Periodic Autoencoder that learns periodic motion features from large unstructured datasets in an unsupervised manner
- Decomposition of movement into multiple latent channels capturing non-linear periodicity of body segments over time, yielding a multi-dimensional phase space
- A learned phase embedding whose feature distances give a better similarity measure than raw motion space, improving temporal/spatial alignment across synthesis tasks
- Context
- Advances the authors' phase-based motion line, from the Neural State Machine (Starke et al. 2019) and Local Motion Phases (Starke et al. 2020), toward fully unsupervised, learned phase extraction.Builds on: Neural State Machine for Character-Scene Interactions · Local Motion Phases for Learning Multi-Contact Character Movements
- Correctness
- Key idea is that an unsupervised periodic embedding aligns and clusters motion better than the original space; demonstrated to help several synthesis tasks (locomotion, style, dance-from-music, dribbling, pose matching), with the reader caveat that gains depend on dataset richness and the chosen number of phase channels.
- Clarity
- Well written and influential (noted as SIGGRAPH 2022 Best Paper); the concept lands on a first pass while the periodic encoder formulation rewards a second.
- How to read it
- First pass for the phase-manifold idea and why phase distance beats motion-space distance; second pass on the Periodic Autoencoder formulation if you intend to use phases as features in your own synthesis pipeline.
Builds on
Related work
- Character Controllers Using Motion VAEs 2020 / SIGGRAPH
- Dog Code: Human to Quadruped Embodiment Using Shared Codebooks 2024 / MIG
- Motion Graphs 2002 / SIGGRAPH
- Neural Animation Layering for Synthesizing Martial Arts Movements 2021 / SIGGRAPH
Keywords
This page summarises the entry and links to its original source. The archive never hosts or redistributes the publication itself.Show it in the full archive list →