3D Mamba-Based Structure

• 3D Mamba-based structure is a design paradigm that adapts selective state space models to process 3D data by preserving local geometry and topology during serialization.
• These architectures integrate multi-directional scans and local operators, such as convolutions and graph fusion, to maintain spatial continuity while modeling long-range dependencies.
• Empirical studies in areas like pose estimation, LiDAR detection, and medical imaging demonstrate improved efficiency and accuracy compared to traditional Transformer-based models.

Searching arXiv for papers on 3D Mamba-based architectures and structure-aware designs. 3D Mamba-based structure denotes a family of architectures that adapt Mamba or selective State Space Models (SSMs) to inherently three-dimensional data, including skeleton sequences, sparse voxels, point clouds, volumetric medical images, articulated meshes, hyperspectral cubes, and speech-driven 3D face motion. Across these settings, a recurring problem is the mismatch between multidimensional geometry and one-dimensional sequence processing: once a 3D signal is serialized, local topology, neighborhood continuity, or articulated structure can be weakened or destroyed. Recent work therefore treats the surrounding architectural design—not merely the presence of an SSM block—as the decisive factor. The resulting systems typically preserve local 3D structure before or during scanning, construct geometry-aware or topology-aware sequence orderings, and use Mamba’s linear-time selective scan for long-range dependency modeling (Cui et al., 12 Nov 2025, Zheng et al., 17 Nov 2025, Wang et al., 26 Jun 2025, Wang et al., 25 Mar 2025).

1. Conceptual scope and defining problem

A 3D Mamba-based structure is not a single model class but a design pattern for applying Mamba to 3D domains where the input is not naturally a 1D sequence. This includes monocular 3D human pose lifting from 2D joint sequences (Cui et al., 12 Nov 2025, Lu et al., 26 Jul 2025, Zheng et al., 27 May 2025), LiDAR-based 3D object detection with sparse voxels (Zheng et al., 17 Nov 2025, Jin et al., 15 Mar 2025, Ning et al., 23 Feb 2026), 3D medical image segmentation on volumetric MRI or CT (Wu et al., 18 Sep 2025, Cao et al., 2024, Wang et al., 25 Mar 2025), point cloud representation learning and completion (Wang et al., 26 Jun 2025, Li et al., 2024, Li et al., 2024), articulated mesh generation and reconstruction (Yoshiyasu et al., 21 Jul 2025), diffusion-based 3D shape generation over voxelized shapes (Mo, 2024), 3D talking-head mesh motion generation (Jafari et al., 2024), and hyperspectral image classification over 3D spectral-spatial cubes (He et al., 2024).

The main architectural difficulty is repeatedly formulated as a serialization problem. In 3D human pose estimation, prior SSM-based methods often flatten $X\in\mathbb{R}^{T\times V\times C}$ into $X_{st}\in\mathbb{R}^{(TV)\times C}$ or $X_{ts}\in\mathbb{R}^{(VT)\times C}$, which “disrupts the human body's spatial topology and entangles temporal and spatial features” (Cui et al., 12 Nov 2025). In LiDAR detection, axis-aligned scanning within fixed windows fragments objects across window boundaries and weakens multi-scale consistency (Zheng et al., 17 Nov 2025). In point cloud learning, serialization destroys Euclidean adjacency because neighboring points need not remain adjacent in sequence order (Wang et al., 26 Jun 2025), and in mesh modeling arbitrary vertex orderings are inadequate because Mamba is not permutation-invariant or graph-aware (Yoshiyasu et al., 21 Jul 2025).

A plausible unifying interpretation is that 3D Mamba-based structure refers less to the internal SSM equations than to the set of operators that make a 3D signal compatible with selective scan without discarding the priors that matter for the task. This interpretation is consistent across skeletons, voxels, points, meshes, volumes, and spectral-spatial cubes.

2. State-space backbone and recurrent architectural motifs

Most of these models inherit the standard continuous-time SSM

$h'(t)=Ah(t)+Bx(t),\qquad y(t)=Ch(t)$

or its discrete form

$$h_t=\overline{A}h_{t-1}+\overline{B}x_t,\qquad y_t=Ch_t$$

with input-dependent selective parameterization in the Mamba style (Cui et al., 12 Nov 2025, Mo, 2024, Lu et al., 26 Jul 2025, Wang et al., 26 Jun 2025, Jin et al., 15 Mar 2025, He et al., 2024). In this literature, the paper-specific novelty rarely lies in a new state equation. Instead, it lies in the transformation of 3D data into sequences, the use of multiple directional scans, and the addition of structure-preserving operators around the SSM.

Several recurring motifs appear across domains. One is the replacement of self-attention by selective scan inside a residual token-mixing block. SasMamba, for example, uses stacked SasMamba blocks, where each block applies SAS-SSM, residual addition, layer normalization, and an MLP: $$\begin{aligned} X^{'}_l &= X_{sas,l}+ X_{l-1} \ X_l &= MLP(LN(X^{'}_l)) + X^{'}_l. \end{aligned}$$ This preserves the familiar token-mixing plus feed-forward structure while replacing self-attention by SAS-SSM (Cui et al., 12 Nov 2025). MedSegMamba similarly uses VSS3D blocks in which LayerNorm, linear projection, depth-wise convolution, SiLU, SS3D, LayerNorm, linear projection, and an MLP are wrapped in residual submodules (Cao et al., 2024). Pamba uses a ConvMamba block with local sparse convolution followed by bidirectional Mamba and an MLP in pre-norm residual form (Li et al., 2024).

A second motif is bidirectionality or multi-directionality. SasMamba explicitly sums temporal-forward, temporal-backward, spatial-forward, and spatial-backward SSM streams (Cui et al., 12 Nov 2025). Pamba reports gains from bidirectional Mamba with shared parameters over unidirectional scanning (Li et al., 2024). MedSegMamba constructs 48 traversal paths through six axis transpositions and eight sequence variants per orientation, with each SS3D module processing one group of eight sequences (Cao et al., 2024). 3DSS-Mamba builds forward and reverse spectral-priority and spatial-priority scans and shows that parallel spectral-spatial routing is best among the tested routes (He et al., 2024).

A third motif is the retention of local operators around the SSM. This is especially explicit in medical imaging and point clouds. HybridMamba argues that overemphasizing global context can compromise local structural information, leading to boundary ambiguity and regional distortion; it therefore combines Slice-oriented Mamba and Local-oriented Mamba with an FFT Gated Mechanism (Wu et al., 18 Sep 2025). The volumetric segmentation analysis in "A Comprehensive Analysis of Mamba for 3D Volumetric Medical Image Segmentation" finds that replacing 1D depthwise convolution by 3D depthwise convolution inside Mamba is decisive, improving average Dice from 85.53 to 87.45 with only slight increases in parameters and FLOPs (Wang et al., 25 Mar 2025). This suggests that local 3D inductive bias remains necessary even when linear-time global modeling is available.

3. Structure preservation before, during, and after scanning

The most characteristic property of 3D Mamba-based structure is explicit preservation of geometry or topology before or during sequence construction. This takes different forms depending on the domain.

In skeleton-based pose lifting, SasMamba introduces a structure-aware spatiotemporal convolution operating directly on the $(T,V)$ grid before scan construction. Given $X\in\mathbb{R}^{T\times V\times C}$, the model predicts offsets $\Delta p\in\mathbb{R}^{T\times V\times 2}$, samples neighboring features with bilinear interpolation, and fuses them with a local convolution branch: $X_{struct}=Conv2d(X) + \sum_{1}^{K^2} {W}_k {X'_{k}.$ This preserves token identity as a joint at a specific frame, without using an explicit skeleton adjacency matrix (Cui et al., 12 Nov 2025). SAMA uses a different mechanism: a learnable, degree-normalized adjacency matrix $X_{st}\in\mathbb{R}^{(TV)\times C}$0 is used for both feature-level and state-level structure fusion,

$X_{st}\in\mathbb{R}^{(TV)\times C}$1

so that joint topology enters directly into state propagation rather than being approximated by scan order alone (Lu et al., 26 Jul 2025). Mamba-Driven Topology Fusion adds a Bone Aware Module and inserts forward and backward GCNs inside Vision Mamba: $X_{st}\in\mathbb{R}^{(TV)\times C}$2 embedding skeletal adjacency in the local preprocessing path of Mamba (Zheng et al., 27 May 2025).

In LiDAR detection, structure preservation often centers on windowing, positional injection, and space-filling curves. WinMamba adds a learnable positional embedding $X_{st}\in\mathbb{R}^{(TV)\times C}$3 from voxel coordinates, uses original and shifted sparse windows through Window Shift Fusion, and adapts window size across scales through Window-Scale Adaptation (Zheng et al., 17 Nov 2025). UniMamba inserts a 3D submanifold convolution before serialization as Spatial Locality Modeling,

$X_{st}\in\mathbb{R}^{(TV)\times C}$4

then uses complementary Z-order serialization and splits channels into local and global groups for Mamba encoding (Jin et al., 15 Mar 2025). Fore-Mamba3D restricts the encoded sequence to sampled foreground voxels, serializes them by rotated Hilbert curves, then repairs the resulting context loss with regional and state-space fusion (Ning et al., 23 Feb 2026).

In point cloud learning, StruMamba3D makes the hidden states themselves spatially meaningful. It defines spatial state anchors by centroids

$X_{st}\in\mathbb{R}^{(TV)\times C}$5

and makes selective parameters state-wise and geometry-conditioned: $X_{st}\in\mathbb{R}^{(TV)\times C}$6 This gives the state space explicit regional semantics (Wang et al., 26 Jun 2025). Pamba instead relies on multi-path space-filling-curve serialization and sparse convolutional local aggregation (Li et al., 2024). 3DMambaComplete uses FPS and EdgeConv to construct sampled point tokens before Mamba, then introduces HyperPoints and HyperPoint Spread to preserve geometric support and spatial coverage (Li et al., 2024).

In articulated meshes, MeshMamba’s key preservation mechanism is serialization itself. Vertices are sorted either lexicographically by template coordinates or by DensePose body-part annotations and UV coordinates so that token order respects articulated structure (Yoshiyasu et al., 21 Jul 2025). In volumetric medical segmentation, MedSegMamba uses axis transpositions and multiple sequence reversals in SS3D (Cao et al., 2024), while the broader volumetric analysis concludes that complex scanning is often unnecessary and that simple scanning frequently suffices, with Tri-scan offering additional gains in the most challenging cases (Wang et al., 25 Mar 2025).

4. Domain-specific realizations

The term spans several distinct 3D regimes, each emphasizing a different structural prior.

Skeletal sequences and human pose

Human pose lifting is one of the clearest cases because the structured object is explicit. SasMamba is a lightweight sequence-to-sequence architecture for monocular 3D human pose estimation from 2D keypoint sequences. It uses $X_{st}\in\mathbb{R}^{(TV)\times C}$7 blocks and hidden dimension $X_{st}\in\mathbb{R}^{(TV)\times C}$8 in the base model, predicts $X_{st}\in\mathbb{R}^{(TV)\times C}$9, and trains with

$X_{ts}\in\mathbb{R}^{(VT)\times C}$0

with $X_{ts}\in\mathbb{R}^{(VT)\times C}$1 and $X_{ts}\in\mathbb{R}^{(VT)\times C}$2 (Cui et al., 12 Nov 2025). SAMA extends structural modeling by injecting graph-like fusion into Mamba2/SSD and making the temporal timescale motion-adaptive through

$X_{ts}\in\mathbb{R}^{(VT)\times C}$3

which the paper interprets as allowing larger timescales for active limbs and smaller ones for stable trunk joints (Lu et al., 26 Jul 2025). Mamba-Driven Topology Fusion further combines spherical-coordinate bone vectors, graph-enhanced Mamba, and alternating spatial-temporal refinement (Zheng et al., 27 May 2025).

Sparse voxels and 3D object detection

In autonomous driving, the dominant concern is efficient long-range modeling over sparse voxel grids. WinMamba builds a 3D backbone with WinMamba layers, Window Shift Fusion, and Adaptive Window Fusion, and reports that on KITTI it improves overall mAP from 70.8 to 73.7 in Table 1, while a later table reports 73.5 (Zheng et al., 17 Nov 2025). UniMamba emphasizes unified spatial-channel representation learning with local and global sequential aggregation and reports 70.2 mAP on nuScenes (Jin et al., 15 Mar 2025). Fore-Mamba3D shifts the emphasis from whole-scene encoding to foreground-only encoding with repair mechanisms, and reports 70.1 mAP and 74.0 NDS on nuScenes test (Ning et al., 23 Feb 2026).

Volumetric medical imaging

HybridMamba is a hierarchical 3D segmentation network built on SegMamba, with S-LMamba blocks and an FFT Gated Mechanism inserted in the encoder (Wu et al., 18 Sep 2025). MedSegMamba is a 3D patch-based hybrid CNN-Mamba model using a U-Net-like encoder-decoder and a bottleneck of nine VSS3D blocks; it reports DSC 0.88383, VS 0.97076, and ASSD 0.33604 on subcortical segmentation, and approximately 20% fewer parameters than SegMambaBot (Cao et al., 2024). The broader analysis paper proposes UlikeMamba, multi-scale Mamba blocks including MSv4, and Tri-scan, and concludes that Mamba can replace Transformers, improve multi-scale representation learning, and often does not require complex scanning (Wang et al., 25 Mar 2025).

Point clouds, shape generation, and completion

StruMamba3D proposes structural SSM blocks for self-supervised point cloud learning and reports 95.1% accuracy on ModelNet40 and 92.75% on the hardest split of ScanObjectNN without voting (Wang et al., 26 Jun 2025). Pamba processes whole-scene point clouds as serialized sequences for semantic segmentation and reports 77.6 mIoU on ScanNet v2, 36.3 on ScanNet200, and 80.3 on nuScenes (Li et al., 2024). 3DMambaComplete uses Mamba encoder-decoder blocks, HyperPoints, and local deformation for completion, reporting 6.907 $X_{ts}\in\mathbb{R}^{(VT)\times C}$4 on PCN (Li et al., 2024). DiM-3D patchifies a $X_{ts}\in\mathbb{R}^{(VT)\times C}$5 voxel tensor and uses bidirectional Mamba blocks inside diffusion, reporting lower GFLOPs than DiT-3D-XL/2 across multiple resolutions and out-of-memory behavior for DiT at $X_{ts}\in\mathbb{R}^{(VT)\times C}$6 while DiM-3D-XL/2 still runs at 18798.04 GFLOPs (Mo, 2024).

Articulated meshes and 3D facial motion

MeshMamba treats dense articulated meshes as serialized vertex sequences and applies Mamba to up to 10,475 vertices. It reports that on an A100 GPU, generating a mesh with 10,475 vertices using 250 DDIM steps takes about 4.5 s with Mamba versus about 28.1 s with a Transformer using FlashAttention, and about 1 s with 50 DDIM steps for reasonable quality (Yoshiyasu et al., 21 Jul 2025). JambaTalk uses a hybrid Transformer–Mamba decoder for speech-driven 3D talking-head generation, with three Mamba layers on both sides of a Transformer layer, and reports LVE 1.7956 and FDD 0.2167 on VOCASET (Jafari et al., 2024).

Spectral-spatial cubes

3DSS-Mamba demonstrates that the same structural principle extends beyond geometry in the usual sense. It treats hyperspectral patches as 3D spectral-spatial cubes, creates tokens $X_{ts}\in\mathbb{R}^{(VT)\times C}$7, and performs 3D spectral-spatial selective scanning. It reports that parallel spectral-spatial scanning performs best and that the final model achieves OA 98.48 on Pavia University, 95.82 on Indian Pines, and 98.37 on Houston 2013 (He et al., 2024).

5. Efficiency, scaling, and empirical trade-offs

A defining attraction of Mamba in 3D is linear sequence modeling. The efficiency claims are concrete in several papers, though they are task-specific and architecture-dependent.

For pose lifting on Human3.6M with $X_{ts}\in\mathbb{R}^{(VT)\times C}$8, SasMamba reports 0.64M parameters, 1.3G MACs, and 5M MACs/frame, while SasMamba-large reports 4.1M parameters, 8.56G MACs, and 35M MACs/frame. These are markedly smaller than Transformer baselines such as MixSTE at 33.6M parameters and 139.0G MACs, MotionBERT at 42.3M and 174.8G, and KTPFormer at 33.7M and 69.5G (Cui et al., 12 Nov 2025). The same paper reports Human3.6M base-model accuracy of P1 = 41.48 mm, P2 = 34.84 mm, and P1$X_{ts}\in\mathbb{R}^{(VT)\times C}$9=21.44 mm with only 0.64M parameters (Cui et al., 12 Nov 2025).

For LiDAR detection, UniMamba reports on Waymo a backbone cost of 61.9 GFLOPs, 1.6M params, and 75.40/73.61 L2 mAP/mAPH, compared with DSVT at 110.2 GFLOPs, 2.7M params, and 74.00/72.10 (Jin et al., 15 Mar 2025). Fore-Mamba3D reports that with $h'(t)=Ah(t)+Bx(t),\qquad y(t)=Ch(t)$0, compared with LION, FLOPs are reduced by 43.7% and FPS is increased by 23.9% (Ning et al., 23 Feb 2026). WinMamba reports improvements over LION-Mamba on KITTI and Waymo while preserving the basic efficiency rationale of local windows plus linear-time Mamba (Zheng et al., 17 Nov 2025).

For point-cloud self-supervision, StruMamba3D reports 15.8M parameters and 4.0 GFLOPs, close to Mamba3D’s 16.9M/3.9G and far lighter than PCM at 34.2M/45.0G (Wang et al., 26 Jun 2025). For whole-scene segmentation, Pamba is memory efficient but not the smallest in parameter count: 82.2M parameters, 296 ms training latency, 5.2G training memory, 183 ms inference latency, and 4.8G inference memory on ScanNet200; its authors attribute much of the parameter count to sparse convolution rather than Mamba (Li et al., 2024).

In 3D medical imaging, the comparison is often against Transformer bottlenecks rather than pure efficiency baselines. MedSegMamba reports significantly improved ASSD over SegMambaBot with approximately 20% fewer parameters (Cao et al., 2024). The volumetric analysis paper reports that UlikeTrans_vanilla is out-of-memory, whereas UlikeMamba_3d achieves average Dice 87.45 with 24.30M parameters and 46.03G FLOPs, compared with UlikeTrans_SRA at average Dice 85.97 with 45.05M parameters and 64.47G FLOPs (Wang et al., 25 Mar 2025).

These results suggest that the efficiency advantage of Mamba is real but conditional. A plausible implication is that the best trade-off is obtained not by removing all local operators, but by reducing the cost of long-range mixing while keeping lightweight local structure-preserving components.

6. Open questions, recurring limitations, and converging design principles

Several limitations recur across the literature. First, many methods still depend on serialization quality. Point-cloud and mesh papers explicitly note that sequence order remains a sensitive variable (Wang et al., 26 Jun 2025, Yoshiyasu et al., 21 Jul 2025), and 3DMambaComplete leaves point-ordering details under-specified (Li et al., 2024). Second, several papers acknowledge that Mamba alone is insufficient for local structure; this is why local convolution, graph convolution, dynamic sampling, or windowing is repeatedly reintroduced (Cui et al., 12 Nov 2025, Wu et al., 18 Sep 2025, Li et al., 2024, Wang et al., 25 Mar 2025). Third, many contributions are architectural rather than theoretical: they modify tokenization, scan construction, or state fusion rather than proposing new SSM mathematics (Zheng et al., 17 Nov 2025, Jin et al., 15 Mar 2025, Wang et al., 25 Mar 2025).

Task-specific failure modes also remain. SasMamba’s supplementary material notes failure cases on extreme poses such as diving, gymnastics, skiing, and aerial rotations, especially under severe 2D detection degradation, rapid short-term motion changes, and top-down views with strong occlusion (Cui et al., 12 Nov 2025). MeshMamba assumes fixed topology and known correspondences, and does not address loose garments or topology change (Yoshiyasu et al., 21 Jul 2025). JambaTalk notes that perceptual quality still lags behind ground truth and that additional modalities such as gestures, gaze, and emotion remain future directions (Jafari et al., 2024). The volumetric analysis paper notes that its claims are empirical and that future work may explore a fully convolution-free 3D Mamba model (Wang et al., 25 Mar 2025).

Across domains, however, several design principles recur with unusual consistency.

First, local structure should be preserved before or within scanning. Examples include SA-Conv in SasMamba (Cui et al., 12 Nov 2025), 3D DWConv in volumetric segmentation (Wang et al., 25 Mar 2025), sparse convolution in UniMamba (Jin et al., 15 Mar 2025), and lightweight convolution over spatial states in StruMamba3D (Wang et al., 26 Jun 2025).

Second, a 3D signal should rarely be reduced to a single naive scan order. Multi-directionality appears as forward/backward temporal and spatial scans (Cui et al., 12 Nov 2025), shifted and axial windows (Zheng et al., 17 Nov 2025), multiple space-filling curves (Li et al., 2024), complementary X/Y Z-order traversals (Jin et al., 15 Mar 2025), rotated Hilbert scans (Ning et al., 23 Feb 2026), or axis-aligned Tri-scan (Wang et al., 25 Mar 2025).

Third, state-space modeling benefits from explicit structural priors. These priors take the form of learned adjacency matrices over joints (Lu et al., 26 Jul 2025), spherical-coordinate bone vectors (Zheng et al., 27 May 2025), spatial state anchors in point clouds (Wang et al., 26 Jun 2025), window-scale adaptation in voxel pyramids (Zheng et al., 17 Nov 2025), or semantic-assisted state regrouping (Ning et al., 23 Feb 2026).

Fourth, multi-scale representation is often treated as essential rather than optional. This is explicit in stride-based scan construction for pose (Cui et al., 12 Nov 2025), adaptive windows in 3D detection (Zheng et al., 17 Nov 2025), multi-scale Mamba blocks in volumetric segmentation (Wang et al., 25 Mar 2025), and hierarchical encoder-decoder designs across medical imaging, LiDAR detection, and mesh reconstruction (Wu et al., 18 Sep 2025, Cao et al., 2024, Yoshiyasu et al., 21 Jul 2025).

Taken together, these works define 3D Mamba-based structure as a structured sequence-modeling paradigm in which Mamba provides linear-time long-range propagation, while geometry-aware tokenization, topology-aware state interaction, and local 3D operators prevent the collapse of spatial meaning under serialization. This suggests that the decisive question for future systems is not whether Mamba can be used in 3D, but how much 3D structure must be encoded into the sequence construction and state evolution for a given domain.