Research map / 3D vision / Human pose estimation
Existence Is Chaos: Enhancing 3D Human Motion Prediction with Uncertainty Consideration
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 a lightweight encoder-decoder for 3D human motion prediction with a self-attention graph generate block (SAGGB) and an Adaptive-Salient Loss that learns per-frame uncertainty weights and emphasizes the first predicted frame.
Problem
3D human motion prediction from past poses, where standard L2 training treats all future frames equally and forces predictions to match a single recorded future despite inherent uncertainty.
Approach
Combines a data-driven, sample-specific self-attention graph encoder with a loss that learns per-frame uncertainty weights (adaptive loss, homoscedastic-style) plus an extra weighted penalty on the first predicted frame (salient loss).
Key results
- On H3.6M short-term (80-400ms) averages, the method reports 9.1/20.9/45.9/57.4 MPJPE vs PGBIG 10.3/22.7/47.4/58.5.
- Jitter on H3.6M 0-1000ms is 111.82 vs 161.20 for PGBIG and 671.76 for STSGCN.
- Uses 0.55M parameters vs 1.74M for PGBIG and 2.55M for LTD, with 8.09ms inference time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed uncertainty-aware method instead of PGBIG | H3.6M motion prediction · H3.6M average, 80ms | MPJPE: 10.3 -> 9.1 | Helped |
| use the proposed uncertainty-aware method instead of PGBIG | H3.6M motion prediction · H3.6M average, 400ms | MPJPE: 58.5 -> 57.4 | Helped |
| use the proposed uncertainty-aware method instead of PGBIG | H3.6M motion prediction · H3.6M average, 560ms | MPJPE: 76.9 -> 77.5 | Hurt |
| use the proposed uncertainty-aware method instead of SPGSN | H3.6M motion prediction · H3.6M average, 1000ms | MPJPE: 109.6 -> 110.4 | Hurt |
| use the proposed uncertainty-aware method instead of STSGCN | H3.6M motion prediction · H3.6M, STSGCN evaluation, 400ms | MPJPE: 38.3 -> 30.2 | Helped |
| use the proposed uncertainty-aware method instead of PGBIG | CMU Mocap motion prediction · CMU Mocap average, 400ms | MPJPE: 36.6 -> 36.4 | Helped |
| use the proposed uncertainty-aware method instead of SPGSN | CMU Mocap motion prediction · CMU Mocap average, 1000ms | MPJPE: 77.8 -> 82.1 | Hurt |
| use the proposed uncertainty-aware method instead of PGBIG* | 3DPW motion prediction · 3DPW average, 100ms | MPJPE: 13.1 -> 12.4 | Helped |
| use the proposed uncertainty-aware method instead of PGBIG* | 3DPW motion prediction · 3DPW average, 1000ms | MPJPE: 109.4 -> 107.7 | Helped |
| use the proposed method instead of PGBIG | H3.6M motion prediction · H3.6M, 0-1000ms | Jitter: 161.20 -> 111.82 | Helped |
| replace GCN with SAGGB (ablation) | H3.6M motion prediction, 6-block model · H3.6M average, 1000ms | MPJPE: 111.8 -> 110.4 | Helped |
| use 6 SAGGB blocks instead of 3 (ablation) | H3.6M motion prediction · H3.6M average, 1000ms | MPJPE: 112.6 -> 110.4 | Helped |
Surprising findings
- Long-term (1000ms) MPJPE on H3.6M is competitive but not best; SPGSN and PGBIG are close or better on several actions.
- Authors argue lower MPJPE alone does not reflect motion quality, as models with similar MPJPE can show very different jitter.
Limitations
- MPJPE may not fully capture motion quality or realism.
- Benchmark datasets were built for pose estimation or action recognition, not motion prediction.
- Long-term prediction is competitive rather than state of the art on several metrics.
Experimental setup
Datasets: Human3.6M CMU Mocap 3DPW
Benchmarks: Human3.6M CMU Mocap 3DPW
Metrics: MPJPE Jitter
Baselines: DMGNN LTD SPGSN PGBIG STSGCN GAGCN
Builds on: LTD PGBIG DMGNN siMLPe
Closest papers by meaning
- Optimal-state Dynamics Estimation for Physics-based Human Motion Capture from Videos (NeurIPS 2024 · similarity 0.87)
- ALIEN: Implicit Neural Representations for Human Motion Prediction under Arbitrary Latency (CVPR 2025 · similarity 0.86)
- Motion Diversification Networks (CVPR 2024 · similarity 0.86)
- Nonisotropic Gaussian Diffusion for Realistic 3D Human Motion Prediction (CVPR 2025 · similarity 0.86)
- SimpliHuMoN: Simplifying Human Motion Prediction (NeurIPS 2026 · similarity 0.86)
- Multimodal Sense-Informed Forecasting of 3D Human Motions (CVPR 2024 · similarity 0.86)
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.