Parts-Mamba Model: SSM Hybrid Architecture

• Parts-Mamba model is a neural architecture that integrates Mamba state-space models with part-aware feature extraction to efficiently capture local and global features.
• It enhances robustness in challenging conditions like occlusion by employing part-specific SSMs and structured fusion mechanisms for spatial and temporal data.
• Empirical results demonstrate significant accuracy gains, reduced inference times, and improved mesh reconstruction fidelity compared to conventional methods.

The term Parts-Mamba Model refers to a family of neural network architectures that integrate state-space models (SSMs)—notably the Mamba SSM—with part-aware or part-level structural processing in complex data domains, predominantly in skeleton-based human action recognition and articulated 3D mesh modeling. These models leverage the efficient long-range dependency modeling capabilities of Mamba SSM alongside explicit inductive biases over spatial or semantic parts, yielding robust performance under challenging conditions such as partial occlusion, large input size, or non-local dependency requirements. Recent instantiations include hybrid GCN-SSM models for skeleton action recognition (Shen et al., 21 Nov 2025), and mesh-sequence SSMs for 3D human mesh generation and reconstruction (Yoshiyasu et al., 21 Jul 2025).

1. Architectural Foundations

Parts-Mamba models are constructed by the combination of SSM (Mamba) blocks with part-aware feature extraction or part-based data serialization.

• Skeleton Action Recognition (Parts-Mamba, (Shen et al., 21 Nov 2025)): The primary architecture is a hybrid stack consisting of:
  • GCN head ("DeGCN") producing initial per-joint features,
  • Mamba "Spatial Fusion" blocks with part-wise and whole-body bi-directional SSMs,
  • Graph modeling and gated fusion combining part- and global-context signals,
  • A Mamba temporal encoder for long-range temporal context integration.

The spatial-temporal block pipeline is:

$$X_\text{raw} \rightarrow \text{DeGCN} \rightarrow X_\text{gcn} \rightarrow \{\text{Parts-Body Scanning} \rightarrow \text{Graph Modeling} \rightarrow \text{Gated Fusion}\} \rightarrow X_\text{spatial} \rightarrow \text{Mamba Temporal Encoder} \rightarrow X_\text{final} \rightarrow \text{Global Pool} \rightarrow \text{Classifier}$$

• 3D Articulated Mesh Generation (MeshMamba, (Yoshiyasu et al., 21 Jul 2025)): The network receives either mesh vertex features or image-extracted mesh queries:
  • Mesh/vertex inputs are serialized using a part-aware or coordinate-based scheme (DensePose-IUV or canonical coordinates),
  • Multiple SSM (Mamba) blocks process the serialized sequence, alternating orderings across layers,
  • Specialized heads for mesh regeneration (diffusion, MambaDiff3D) or regression from image (Mamba-HMR).

2. Part-Aware and Selective State-Space Modeling

Central to Parts-Mamba is the application of SSMs in a part-level or part-aware manner.

• Skeleton Model (Shen et al., 21 Nov 2025): For each anatomical part $\rho$ (e.g., limb subsets), an independent SSM (bidirectional, i.e., both input and reversed order) is applied to features down-projected from the DeGCN head. For partwise features $X_p$:

$$X_p \leftarrow \text{Conv1D}\left(\bigg\Vert_{\rho\in\text{Parts}} [\text{SSM}_\rho(X_p) \oplus \text{SSM}_\rho(X_p^r)]\right)$$

Body-wise features $X_s$ are similarly processed. Elementwise sums $\oplus$ and concatenations $\Vert$ capture both local-part and global context.

• 3D Mesh Model (Yoshiyasu et al., 21 Jul 2025): Vertices are serialized to prioritize anatomical or spatial locality. This sequence is then processed using linear-time SSM scans:

$$h_{t+1} = A(x_t) h_t + B(x_t) x_t\,, \quad y_t = C(x_t) h_t$$

Each layer can process an ordering that places all vertices of a part contiguously, which empirically improves sample quality and learning efficiency.

3. Fusion Mechanisms and Graph Integration

After separate SSM processing, Parts-Mamba models implement structured fusion designed to preserve both local part features and non-local joint/body context.

• In (Shen et al., 21 Nov 2025), spatial part-body fusion is achieved by:
  • Injecting graph topology via adjacency-matrix-driven aggregation ($A_p, A_s$ learnable),
  • Aggregating five streams—part SSM, body SSM, their graph-enhanced versions, and a gated global channel,
  • Combining these through elementwise multiplications and a channel-projection conv:

  $$F_\text{self} = (X_p * X_g) \oplus (X_s * X_g)\,,\quad F_\text{cross} = (X_p' * X_s) \oplus (X_p * X_s')$$

  $\rho$0

• In MeshMamba (Yoshiyasu et al., 21 Jul 2025), fusion is less explicit, but the use of alternating part-based and coordinate-based orderings across layers promotes both local shape and global articulation capture.

4. Occlusion Robustness and Data Masking Strategies

A primary motivation for part-level modeling in skeleton and mesh contexts is robustness to occlusion and missing data.

• Skeleton Occlusion (Shen et al., 21 Nov 2025):

  • Spatial occlusion is simulated by zero-masking all joints in a given anatomical part: $\rho$1.
  • Temporal occlusion is simulated by masking contiguous or random frame blocks, $\rho$2.
  • No explicit imputation: zeros signal missingness, and SSMs learn to propagate information from distant, unoccluded context.
• 3D Mesh Occlusion (Yoshiyasu et al., 21 Jul 2025):
  • The serialization approach ensures that discontinuities due to missing vertices are confined within parts, maintaining local structure.
  • Efficient handling of incomplete shapes stems from the same local-global division.

5. Empirical Performance and Benchmarking

Quantitative results across multiple domains illustrate the impact of the Parts-Mamba design.

Task/Dataset	Method	Accuracy/Metric	Relative Improvement
NTU-60, Part Spatial Occlusion	ST-GCN	64.4%	Baseline
	MSFGCN	80.3%	+16.0
	Parts-Mamba (Shen et al., 21 Nov 2025)	84.4%	+20.0
NTU-120, Cross Setup Occlusion	MSFGCN	73.4%	Baseline
	Parts-Mamba (Shen et al., 21 Nov 2025)	82.7%	+9.3
NTU-60, Temporal Occlusion	2s-PD-GCN	≈64.6%	Baseline
	Parts-Mamba (Shen et al., 21 Nov 2025)	76.8%	+12.2
SURREAL, Mesh Generation 1-NNA↓	DiffSurf	54.4%	Baseline
	MeshMamba (Yoshiyasu et al., 21 Jul 2025)	53.1%	+1.3
UBody, Mesh Regression PA-MVE↓	Best prior	~32mm	Baseline
	MeshMamba (Yoshiyasu et al., 21 Jul 2025)	25.9mm	+6.1mm

Parts-Mamba models achieved up to 13% absolute accuracy gain under heavy occlusion on NTU-60/120 skeleton datasets (Shen et al., 21 Nov 2025). MeshMamba realized significant improvements in generation fidelity (FID 0.32 vs 0.64) and an order-of-magnitude speed-up versus Transformer algorithms (inference on A100: 4.5s for MeshMamba vs 28.1s) with no loss in mesh reconstruction precision (Yoshiyasu et al., 21 Jul 2025).

6. Applications, Limitations, and Interpretation

Parts-Mamba models have been demonstrated in:

• Action recognition with incomplete skeletons: Enhanced robustness under joint occlusion or missing frames due to their ability to aggregate context within parts and globally (Shen et al., 21 Nov 2025).
• Dense human mesh generation and recovery: Scalable to thousands of vertices, with explicit part-aware semantic structure, fast enough for practical deployment (Yoshiyasu et al., 21 Jul 2025).

A key implication is that structured SSMs, when combined with part-specific orderings and part-level input masking, generalize better to data with partial observations—a typical characteristic in real-world vision and sensor networks.

A plausible implication is that these architecture principles could be extended beyond human pose/mesh domains to hierarchical or modular data in other fields, leveraging Mamba SSM’s linear-time complexity on part-structured sequences.

7. Related Work and Evolution

Parts-Mamba models synthesize developments in:

• Graph Convolutional Networks for structured data, integrating local topological priors.
• State-Space Models (Mamba, S4, etc.) for memory-efficient long-sequence modeling, offering linear-time, hardware-friendly scans.
• Part-based ordering and fusion, inspired by anatomical, semantic, or spatial hierarchies.

While TransMamba models in NLP also employ SSM ("Mamba parts") sub-blocks with shared QKV parameters and dynamic branch switching (Li et al., 31 Mar 2025), the term Parts-Mamba itself specifically denotes part-level SSM hybridization in structured spatial domains (Shen et al., 21 Nov 2025, Yoshiyasu et al., 21 Jul 2025). Controversies remain limited, but broader adoption of part-aware SSMs will likely prompt new benchmarks and comparisons in domains with multi-part or occluded signals.