Research map / 3D vision / Human pose estimation
Pose Magic: Efficient and Temporally Consistent Human Pose Estimation with a Hybrid Mamba-GCN Network
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 Pose Magic, an attention-free dual-stream Mamba+GCN network for 2D-to-3D human pose lifting with a bidirectional and a causal variant, reaching 37.5 mm MPJPE on Human3.6M with 74.1% fewer FLOPs than the prior best Transformer.
Problem
3D human pose estimation from 2D pose sequences faces a trade-off between accuracy and computational efficiency in Transformer-based backbones, and existing methods lack temporal smoothness and causal real-time variants.
Approach
Combine a Mamba stream (global forward/backward/independent selective SSM paths) with a GCN stream (skeleton-adjacency spatial graph and K-NN temporal graph) via adaptive softmax fusion, plus a causal masked K-NN adjacency for real-time use.
Key results
- Bidirectional Pose Magic reaches 37.5 mm MPJPE on Human3.6M (T=243), 0.9 mm better than the prior best (38.4 mm) while saving 74.1% FLOPs (40.58G).
- Causal Pose Magic reaches 41.7 mm MPJPE on Human3.6M vs 42.6 mm for the re-implemented causal baseline.
- On MPI-INF-3DHP with T=81, Pose Magic achieves 14.7 mm MPJPE, 98.8 PCK and 87.6 AUC in the bidirectional setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Pose Magic instead of the strongest reported baseline | Pose Magic causal · Human3.6M | MPJPE (mm): 42.6 -> 41.7 | Helped |
| Use Pose Magic instead of the strongest reported baseline | Pose Magic bidirectional · Human3.6M | MPJPE (mm): 38.4 -> 37.5 | Helped |
| Use Pose Magic bidirectional instead of the strongest baseline | Pose Magic · MPI-INF-3DHP bidirectional | MPJPE: 16.2 -> 14.7 | Helped |
| Use Pose Magic bidirectional instead of the strongest baseline | Pose Magic · MPI-INF-3DHP bidirectional | PCK: 99.1 -> 98.8 | Hurt |
| Use Pose Magic bidirectional instead of the strongest baseline | Pose Magic · MPI-INF-3DHP bidirectional | AUC: 88.0 -> 87.6 | Hurt |
| Add the GCN stream to Mamba in parallel (ablation) | Pose Magic · Human3.6M bidirectional | MPJPE: 41.2 -> 37.5 | Helped |
| Add the GCN stream to Mamba in parallel (ablation) | Pose Magic · Human3.6M bidirectional | MPJVE: 1.8 -> 1.6 | Helped |
| Use GCN alone instead of the hybrid Mamba-GCN block (ablation) | Pose Magic · Human3.6M bidirectional | MPJPE: 37.5 -> 54.2 | Hurt |
| Use GCN alone instead of the hybrid Mamba-GCN block (ablation) | Pose Magic · Human3.6M bidirectional | MPJVE: 1.6 -> 4.6 | Hurt |
| Connect GCN then Mamba sequentially instead of parallel streams (ablation) | Pose Magic · Human3.6M bidirectional | MPJPE: 37.5 -> 44.2 | Hurt |
| Connect Mamba then GCN sequentially instead of parallel streams (ablation) | Pose Magic · Human3.6M bidirectional | MPJPE: 37.5 -> 46.8 | Hurt |
| Add temporal positional embedding to spatial positional embedding (ablation) | Pose Magic · Human3.6M bidirectional | MPJPE: 37.5 -> 38.8 | Hurt |
Surprising findings
- Adding temporal positional embedding on top of spatial embedding increased MPJPE by 1.3 mm, which the authors attribute to GCN's non-permutation-equivariance and Mamba's temporal continuity.
Limitations
- Causal variant is evaluated on Human3.6M and MPI-INF-3DHP only with comparisons limited to re-implemented baselines.
- Ablation on positional embedding shows temporal embedding hurts MPJPE, so design choices are dataset-specific.
Experimental setup
Datasets: Human3.6M MPI-INF-3DHP
Benchmarks: Human3.6M MPI-INF-3DHP
Metrics: MPJPE MPJVE ACC-ERR PCK AUC Params FLOPs
Baselines: MotionBERT MotionAGFormer PoseFormer MixSTE STCFormer MHFormer HSTFormer P-STMO GLA-GCN Strided Transformer
Builds on: Mamba MotionBERT MotionAGFormer GCN S4D
Closest papers by meaning
- VIMCAN: Visual-Inertial 3D Human Pose Estimation with Hybrid Mamba-Cross-Attention Network (CVPR 2026 · similarity 0.89)
- PoseMamba: Monocular 3D Human Pose Estimation with Bidirectional Global-Local Spatio-Temporal State Space Model (AAAI 2025 · similarity 0.88)
- StructMamPose: From Sequential Perception to Structural Reasoning for 3D Human Pose Estimation (ICML 2026 · similarity 0.85)
- Non-local Latent Relation Distillation for Self-Adaptive 3D Human Pose Estimation (NeurIPS 2021 · similarity 0.85)
- Deep Reinforcement Learning for Active Human Pose Estimation (AAAI 2020 · similarity 0.85)
- MV-SSM: Multi-View State Space Modeling for 3D Human Pose Estimation (CVPR 2025 · similarity 0.85)
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.