Research map / 3D vision / Human pose estimation
ALIEN: Implicit Neural Representations for Human Motion Prediction under Arbitrary Latency
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.
Proposes ALIEN, a motion implicit neural representation with a Mamba-like linear attention hyper-network, low-rank modulation, and an auxiliary variable-delay reconstruction task, for human motion prediction under arbitrary latency.
Problem
Forecasting future human poses when the latency between observation and action varies unpredictably across samples (network transmission plus algorithm execution), which existing motion predictors ignore.
Approach
Represent motion as a continuous function via a coordinate-based INR whose weights are generated from past observations by a linear-complexity MLLA hyper-network with low-rank modulation; predictions are made by querying arbitrary future temporal coordinates, and an auxiliary variable-delay pose reconstruction task is trained jointly.
Key results
- On Human3.6M under arbitrary latency, ALIEN lowers average MPJPE versus NeRMo by 5.83%, 3.54%, 3.29%, 3.64% and 6.28% at 80, 160, 320, 400 and 600 ms.
- On CMU-MoCap and 3DPW, ALIEN reduces MPJPE by 4.08% and 6.26% at 80 ms relative to the best competing methods.
- Removing the auxiliary reconstruction task on Human3.6M raises 80 ms MPJPE from 9.7 to 10.4 and 400 ms from 58.3 to 60.1 (Table 6).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ALIEN instead of the strongest listed baseline | Human3.6M motion prediction under arbitrary latency · Human3.6M, average, 80ms | MPJPE: 10.3 -> 9.7 | Helped |
| Use ALIEN instead of the strongest listed baseline | Human3.6M motion prediction under arbitrary latency · Human3.6M, average, 600ms | MPJPE: 82.2 -> 79.3 | Helped |
| Use ALIEN instead of the strongest listed baseline | CMU-MoCap motion prediction under arbitrary latency · CMU-MoCap, average, 320ms | MPJPE: 27.86 -> 28.06 | Hurt |
| Use ALIEN instead of the strongest listed baseline | CMU-MoCap motion prediction under arbitrary latency · CMU-MoCap, average, 400ms | MPJPE: 36.15 -> 35.93 | Helped |
| Use ALIEN instead of the strongest listed baseline | 3DPW motion prediction under arbitrary latency · 3DPW, average, 100ms | MPJPE: 15.75 -> 14.83 | Helped |
| Use ALIEN instead of the strongest listed baseline | 3DPW motion prediction under arbitrary latency · 3DPW, average, 800ms | MPJPE: 104.10 -> 103.62 | Helped |
| Add low-rank modulation (ablation) | ALIEN · Human3.6M | RunTime: 38.73 -> 29.10 | Helped |
| Add low-rank modulation (ablation) | ALIEN · Human3.6M, 400ms | MPJPE: 58.2 -> 58.3 | No clear effect |
| Add the auxiliary variable-delay pose reconstruction task (ablation) | ALIEN · Human3.6M, 80ms | MPJPE: 10.4 -> 9.7 | Helped |
| Add the auxiliary variable-delay pose reconstruction task (ablation) | ALIEN · CMU-MoCap, 80ms | MPJPE: 8.2 -> 7.8 | Helped |
| Replace the MLLA hyper-network with TransINR's Transformer hyper-network (ablation) | ALIEN · Human3.6M, 80ms | MPJPE: 9.7 -> 9.8 | Hurt |
| Replace the MLLA hyper-network with GINR's Transformer hyper-network (ablation) | ALIEN · Human3.6M, 80ms | MPJPE: 9.7 -> 10.3 | Hurt |
Surprising findings
- Baselines adapted to arbitrary latency degrade relative to latency-free prediction because they must predict extra irrelevant poses in the expanded window.
- Standard Mamba without the MLLA design did not perform well because the large number of initialization tokens cannot be ordered.
Limitations
- Evaluated only on fixed latency grids (40 ms steps, up to 200 ms) with simulated latency rather than real network traces.
- Baselines were adapted by the authors since no prior work besides NeRMo addresses this task.
- Slightly less accurate than TransINR, trading accuracy for efficiency.
Experimental setup
Datasets: Human3.6M CMU-MoCap 3DPW
Benchmarks: Human3.6M CMU-MoCap 3DPW
Metrics: MPJPE
Baselines: Res-sup LTD MSR-GCN PGBIG SPGSN CIST NeRMo TransINR GINR
Builds on: NeRF TransINR MLLA NeRMo LTD
Closest papers by meaning
- A Neural Temporal Model for Human Motion Prediction (CVPR 2019 · similarity 0.88)
- LAL: Enhancing 3D Human Motion Prediction with Latency-aware Auxiliary Learning (CVPR 2025 · similarity 0.87)
- Existence Is Chaos: Enhancing 3D Human Motion Prediction with Uncertainty Consideration (AAAI 2024 · similarity 0.86)
- SimpliHuMoN: Simplifying Human Motion Prediction (NeurIPS 2026 · similarity 0.85)
- HVIS: A Human-like Vision and Inference System for Human Motion Prediction (AAAI 2025 · similarity 0.85)
- Motion Diversification Networks (CVPR 2024 · similarity 0.84)
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.