Research map / 3D vision / Human pose estimation
PoseMamba: Monocular 3D Human Pose Estimation with Bidirectional Global-Local Spatio-Temporal State Space Model
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 PoseMamba, a purely state-space-model (Mamba-based) lifter for monocular 3D human pose estimation that uses bidirectional global-local spatio-temporal scanning to reach SOTA on Human3.6M and MPI-INF-3DHP with fewer parameters and MACs.
Problem
Transformer-based 3D human pose estimation from monocular video suffers quadratic attention cost and unidirectional, insufficient modeling of spatio-temporal joint correlations.
Approach
Replace self-attention with bidirectional Mamba (S6) blocks that scan spatial and temporal tokens forward and backward, plus a reordered skeleton scan (local ordering merged with the global scan) to improve limb prediction at linear complexity.
Key results
- PoseMamba-X reaches 37.1 mm P1 / 31.5 mm P2 on Human3.6M (2D detections) and 14.8 mm on ground-truth 2D, the best reported in its table, with 26.5M params and 109.9G MACs.
- PoseMamba-B achieves 14.51 mm MPJPE on MPI-INF-3DHP with 81 frames, the best in its table.
- PoseMamba-S (0.86M params) reaches 41.8 mm on Human3.6M in ablation, versus 42.4 mm for bidirectional and 43.0-43.8 mm for unidirectional scans.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add bidirectional global-local spatial-temporal scanning instead of bidirectional scanning without local scan (ablation) | PoseMamba-S · Human3.6M | MPJPE: 42.4 -> 41.8 | Helped |
| replace unidirectional spatial-temporal scan with bidirectional global-local scanning (ablation) | PoseMamba-S · Human3.6M | MPJPE: 43.0 -> 41.8; 43.1 -> 41.8; 43.2 -> 41.8; 43.8 -> 41.8 | Helped |
| add 2D-Loss (ablation) | PoseMamba-S with MPJPE Loss · Human3.6M | MPJPE: 43.7 -> 43.5 | Helped |
| add T-Loss (ablation) | PoseMamba-S with MPJPE Loss · Human3.6M | MPJPE: 43.7 -> 42.1 | Helped |
| add T-Loss to MPJPE Loss + 2D-Loss (ablation) | PoseMamba-S · Human3.6M | MPJPE: 43.5 -> 41.8 | Helped |
| add 2D-Loss to MPJPE Loss + T-Loss (ablation) | PoseMamba-S · Human3.6M | MPJPE: 42.1 -> 41.8 | Helped |
| increase depth from 12 to 16 blocks at 64 channels (ablation) | PoseMamba · Human3.6M | MPJPE: 43.0 -> 42.0 | Helped |
| increase depth from 16 to 20 blocks at 64 channels (ablation) | PoseMamba · Human3.6M | MPJPE: 42.0 -> 41.8 | Helped |
| increase depth from 20 to 24 blocks at 64 channels (ablation) | PoseMamba · Human3.6M | MPJPE: 41.8 -> 41.5 | Helped |
| increase depth from 24 to 32 blocks at 64 channels (ablation) | PoseMamba · Human3.6M | MPJPE: 41.5 -> 41.5 | No clear effect |
| increase depth from 32 to 40 blocks at 64 channels (ablation) | PoseMamba · Human3.6M | MPJPE: 41.5 -> 41.1 | Helped |
| increase model dimension from 32 to 64 channels at 40 blocks (ablation) | PoseMamba · Human3.6M | MPJPE: 43.1 -> 41.1 | Helped |
Surprising findings
- Unidirectional Mamba scans underperform bidirectional ones, and relying solely on Mamba did not reach optimal pose accuracy.
Limitations
- Local-scan reordering is a heuristic skeleton ordering rather than a learned one
- Reported MPI-INF-3DHP results use only 27 and 81 frames, not the 243-frame setting
- Code and models are only promised ('will be released'), not yet available
Experimental setup
Datasets: Human3.6M MPI-INF-3DHP
Benchmarks: Human3.6M MPI-INF-3DHP
Metrics: MPJPE P-MPJPE Params MACs
Baselines: MHFormer MixSTE P-STMO STCFormer PoseFormerV2 MotionBERT MotionAGFormer-L HSTFormer
Builds on: Mamba VMamba Vision Mamba MotionBERT MotionAGFormer
Closest papers by meaning
- MV-SSM: Multi-View State Space Modeling for 3D Human Pose Estimation (CVPR 2025 · similarity 0.92)
- Sequential Joint Dependency Aware Human Pose Estimation with State Space Model (AAAI 2025 · similarity 0.90)
- Global-to-Local Modeling for Video-Based 3D Human Pose and Shape Estimation (CVPR 2023 · similarity 0.89)
- Pose Magic: Efficient and Temporally Consistent Human Pose Estimation with a Hybrid Mamba-GCN Network (AAAI 2025 · similarity 0.88)
- Video-Based Human Pose Regression via Decoupled Space-Time Aggregation (CVPR 2024 · similarity 0.87)
- SVTformer: Spatial-View-Temporal Transformer for Multi-View 3D Human Pose Estimation (AAAI 2025 · 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.