Research map / Label-efficient and robust learning / Recognition and long-tail
MoML: Online Meta Adaptation for 3D Human Motion Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces online meta adaptation for 3D human motion prediction, adding MoAdapter blocks adapted with a few MAML-style inner gradient steps on recent prediction errors (and a closed-form ridge variant, Fast-MoML) on top of existing offline predictors.
Problem
Human motion predictors are trained offline with fixed parameters, which is suboptimal for streaming motion data whose distribution changes over time (concept drift).
Approach
Meta-learn a generic predictor together with lightweight adaptive adapter parameters (MoAdapters); at inference, adapt only the MoAdapters with a few gradient steps on the most recent prediction error, and use a closed-form ridge-regression last layer in the fast variant.
Key results
- On Human3.6M, MoML variants lower MPJPE for LTD, MotionMixer and SPGSN in most of the 5 typical activities; e.g., SPGSN-GC averaged 57.4 vs baseline 58.2 in the 60-sub-task streaming average at 400ms.
- On CMU-Mocap average at 400ms, SPGSN improves from 37.0 to 35.1 (FC) and 35.2 (GC); on 3DPW at 400ms, SPGSN goes from 64.5 to 61.2 (FC).
- Vanilla MAML over the full network gives smaller gains than MoML (e.g., SPGSN 58.22 baseline, 57.34 with MAML, 56.59 with FC+MAML at 400ms over streaming).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FC-MoAdapters with MoML | SPGSN · CMU-Mocap average, 400ms | MPJPE: 37.0 -> 35.1 | Helped |
| Add GC-MoAdapters with MoML | SPGSN · 3DPW average, 400ms | MPJPE: 64.5 -> 61.6 | Helped |
| Add FC-MoAdapters with MoML | SPGSN · Human3.6M discussion, 400ms | MPJPE: 67.1 -> 65.2 | Helped |
| Add FC-MoAdapters with MoML | SPGSN · Human3.6M walking, 400ms | MPJPE: 41.5 -> 40.0 | Helped |
| Add FC-MoAdapters with MoML | SPGSN · Human3.6M eating, 400ms | MPJPE: 37.9 -> 36.8 | Helped |
| Use a last-layer closed-form adapter instead of FC/GC MoAdapters (ablation) | LTD, MotionMixer, and SPGSN · Human3.6M, 400ms testpoint | MPJPE: 61.1; 61.7; 57.9 | Mixed |
| Apply vanilla MAML to all model parameters (ablation) | LTD · Human3.6M, 400ms testpoint | MPJPE: 63.52 -> 61.98 | Helped |
| Use FC-MoAdapters and apply MAML to the whole network (ablation) | LTD · Human3.6M, 400ms testpoint | MPJPE: 63.52 -> 62.51 | Helped |
| Use GC-MoAdapters and apply MAML to the whole network (ablation) | LTD · Human3.6M, 400ms testpoint | MPJPE: 63.52 -> 61.90 | Helped |
| Use gradient-based last-layer optimization instead of the closed-form solution (ablation) | LTD, MotionMixer, and SPGSN · Human3.6M, 400ms testpoint | MPJPE: 62.37; 58.28; 57.58 | Mixed |
Surprising findings
- Updating parameters on static motions (e.g., sitting, waiting) can be unnecessary and sometimes hurts accuracy.
- Full-network vanilla MAML at inference brings only limited gains and is less stable than adapting a small set of adapter parameters.
Limitations
- Fast-MoML's single-layer embedding is limited when facing varied motions, reducing its adaptability.
- Gains are not uniform; some activities (static ones) show failures.
- Evaluated only on three motion datasets and three baselines.
Experimental setup
Datasets: Human3.6M CMU-Mocap 3DPW
Benchmarks: Human3.6M CMU-Mocap 3DPW
Metrics: MPJPE
Baselines: Res. sup DMGNN MSR LTD SPGSN MotionMixer MAML
Builds on: MAML Reptile LTD SPGSN MotionMixer
Closest papers by meaning
- SimpliHuMoN: Simplifying Human Motion Prediction (NeurIPS 2026 · similarity 0.86)
- LAL: Enhancing 3D Human Motion Prediction with Latency-aware Auxiliary Learning (CVPR 2025 · similarity 0.85)
- Harmonizing Stochasticity and Determinism: Scene-responsive Diverse Human Motion Prediction (NeurIPS 2024 · similarity 0.84)
- Motion Diversification Networks (CVPR 2024 · similarity 0.84)
- Existence Is Chaos: Enhancing 3D Human Motion Prediction with Uncertainty Consideration (AAAI 2024 · similarity 0.83)
- SingularTrajectory: Universal Trajectory Predictor Using Diffusion Model (CVPR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.