Two-Stream Mamba Pyramid Network

• The paper demonstrates a novel dual-stream design where a visual TES stream and an audio-visual PCS stream are modeled through a multi-scale Mamba temporal pyramid for figure skating assessment.
• The architecture integrates Mamba modules for long-range temporal encoding and employs a literal pyramid structure to capture hierarchical features across multiple scales.
• Variants in adjacent applications, such as MRI reconstruction and hyperspectral segmentation, confirm that combining explicit stream separation with pyramid-based fusion enhances task-specific feature extraction.

to=arxiv_search 彩神争霸提现ితం code: {"query":"arXiv (Wang et al., 22 Aug 2025) Learning Long-Range Action Representation by Two-Stream Mamba Pyramid Network for Figure Skating Assessment", "max_results": 5} to=arxiv_search 天天爱彩票中奖se code: {"query":"arXiv (Chen et al., 7 Aug 2025) HiFi-Mamba Dual-Stream W-Laplacian Enhanced Mamba for High-Fidelity MRI Reconstruction", "max_results": 5} to=arxiv_search 天天中彩票和 code: {"query":"arXiv (Lin et al., 24 Feb 2025) MDN Mamba-Driven Dualstream Network For Medical Hyperspectral Image Segmentation", "max_results": 5} to=arxiv_search 天天彩票app code: {"query":"arXiv (Zheng et al., 20 Mar 2026) CS-MUNet A Channel-Spatial Dual-Stream Mamba Network for Multi-Organ Segmentation", "max_results": 5} to=arxiv_search 天天中彩票大神推荐 code: {"query":"arXiv (Wang et al., 2024) PyramidMamba Rethinking Pyramid Feature Fusion with Selective Space State Model for Semantic Segmentation of Remote Sensing Imagery", "max_results": 5} to=arxiv_search 微信公众号天天中彩票 code: {"query":"arXiv (Anand et al., 8 Dec 2025) UltrasODM A Dual Stream Optical Flow Mamba Network for 3D Freehand Ultrasound Reconstruction", "max_results": 5} A Two-Stream Mamba Pyramid Network is an architectural pattern that combines dual-stream decomposition, Mamba-based state-space sequence modeling, and pyramid-style multi-scale processing. The designation is used explicitly in figure skating action quality assessment, where a visual-feature-based Technical Element Score (TES) stream is separated from an audio-visual-feature-based Program Component Score (PCS) stream and both are modeled through a multi-scale Mamba temporal pyramid (Wang et al., 22 Aug 2025). In adjacent literature, the label applies unevenly: some models are explicitly two-stream but only weakly pyramid-like, while others are strongly pyramid-like but not explicitly two-stream (Chen et al., 7 Aug 2025, Lin et al., 24 Feb 2025, Zheng et al., 20 Mar 2026, Wang et al., 2024).

1. Scope and architectural taxonomy

The phrase covers a narrow literal usage and a broader family of related designs. The narrow usage refers to the figure-skating model that is named a “two-stream Mamba pyramid network” and aligns stream structure with judging criteria: TES is evaluated from visual evidence only, whereas PCS is evaluated from visual and audio evidence (Wang et al., 22 Aug 2025). The broader usage includes architectures that combine two complementary representational paths with Mamba modules and some form of hierarchical or pyramid processing, even when the papers themselves avoid the exact label.

Paper	Two-stream status	Pyramid status
"Learning Long-Range Action Representation by Two-Stream Mamba Pyramid Network for Figure Skating Assessment" (Wang et al., 22 Aug 2025)	Explicit visual TES stream and audio-visual PCS stream	Explicit multi-scale Mamba temporal pyramid
"HiFi-Mamba: Dual-Stream W-Laplacian Enhanced Mamba for High-Fidelity MRI Reconstruction" (Chen et al., 7 Aug 2025)	Explicit low-/high-frequency streams	Single-level WL decomposition; not a true multi-level pyramid
"MDN: Mamba-Driven Dualstream Network For Medical Hyperspectral Image Segmentation" (Lin et al., 24 Feb 2025)	Explicit spatial and spectral streams	U-shaped hierarchical encoder-decoder, but no explicit feature pyramid
"CS-MUNet: A Channel-Spatial Dual-Stream Mamba Network for Multi-Organ Segmentation" (Zheng et al., 20 Mar 2026)	Explicit channel and spatial/boundary streams	U-shaped multiscale hierarchy and multi-kernel fusion, but no explicit FPN head
"PyramidMamba: Rethinking Pyramid Feature Fusion with Selective Space State Model for Semantic Segmentation of Remote Sensing Imagery" (Wang et al., 2024)	Single main semantic path plus low-level detail path	Explicit dense pyramid pooling plus Pyramid Fusion Mamba

This distribution suggests that “Two-Stream Mamba Pyramid Network” is most precise when all three components are present simultaneously: explicit stream separation, Mamba-based long-range modeling, and a genuine multi-level pyramid. In many later or neighboring papers, only two of the three conditions are fully satisfied.

2. Core design principles

Across the literature, the “two-stream” dimension is defined by task-specific complementarity rather than by a fixed modality split. In figure skating assessment, the split follows the rubric: a visual TES stream for technical elements and an audio-visual PCS stream for artistic and musical interpretation (Wang et al., 22 Aug 2025). In MRI reconstruction, the split is spectral: a low-frequency stream supports long-range global anatomical modeling, while a high-frequency stream preserves edges, textures, and other fine details (Chen et al., 7 Aug 2025). In medical hyperspectral segmentation, the streams are spatial and spectral (Lin et al., 24 Feb 2025). In abdominal multi-organ segmentation, the streams are orthogonal sequence spaces: a channel-semantic stream and a spatial/boundary-aware stream (Zheng et al., 20 Mar 2026).

The “Mamba” component is likewise inserted in different places depending on the target dependency structure. Some models use Mamba as the temporal backbone at every pyramid level, as in the figure-skating network’s Temporal Hierarchical Feature Encoder and Mamba Down Sampling blocks (Wang et al., 22 Aug 2025). Others use Mamba to modulate or fuse multi-scale features after pyramid construction, as in PyramidMamba’s Pyramid Fusion Mamba decoder (Wang et al., 2024). Still others modify Mamba’s state-space parameters directly: HiFi-Mamba adds high-frequency conditioning terms into the $B$ and $C$ branches of the selective state-space model (Chen et al., 7 Aug 2025), whereas CS-MUNet injects boundary posterior information into $\Delta$ and $B$ during boundary-aware spatial fusion (Zheng et al., 20 Mar 2026).

The “pyramid” dimension is the least uniform. In the figure-skating model it is a literal temporal pyramid with six scales, $T, T/2, T/4, T/8, T/16, T/32$ (Wang et al., 22 Aug 2025). In PyramidMamba it is a dense spatial pyramid pooling decoder built from a single high-level feature map (Wang et al., 2024). In MDN and CS-MUNet it is better described as a U-shaped hierarchical encoder-decoder with multiresolution skip fusion rather than a formal feature pyramid network (Lin et al., 24 Feb 2025, Zheng et al., 20 Mar 2026). In HiFi-Mamba, the “pyramid” label is only loosely justified by Laplacian-pyramid-inspired frequency residual splitting, not by a true multi-level downsampling and upsampling hierarchy (Chen et al., 7 Aug 2025).

3. Canonical instantiation in figure skating assessment

The most literal and complete realization appears in "Learning Long-Range Action Representation by Two-Stream Mamba Pyramid Network for Figure Skating Assessment" (Wang et al., 22 Aug 2025). The method is designed around the official decomposition of figure-skating scores into TES and PCS. Each technical element is formalized as

$P_i = (s_i, e_i, c_i, r_i),$

where $s_i$ and $e_i$ are the start and end times, $c_i$ is the action category, and $r_i$ is the element score. The overall PCS is a scalar $C$0.

The pipeline begins with video and audio. Video features are extracted using I3D, audio features using VGGish, and both are projected to a common embedding space,

$C$1

A Temporal Hierarchical Feature Encoder preserves temporal resolution while refining sequence structure. Its Temporal Embedding Module applies stacked 1D convolutions,

$C$2

Its Temporal Refinement Module then applies masked Mamba blocks,

$C$3

After THFE, a six-level temporal pyramid is built by Mamba Down Sampling. The visual pyramid serves the TES stream. At each level $C$4, the TES head produces class logits, temporal offsets, and per-element scores from video features $C$5: $C$6

$C$7

The regression ranges are $C$8, reflecting the fact that action durations range from $C$9 to $\Delta$0 seconds. Inference applies soft-NMS with IoU threshold $\Delta$1, keeps the top $\Delta$2 segments for short programs or top $\Delta$3 for free skating, and computes

$\Delta$4

The PCS stream is asymmetric. It uses both visual and audio features, but fusion is designed so that audio does not contaminate TES. Multi-level Cross Attention Fusion operates at each pyramid level $\Delta$5, with video features as query and audio features as key and value: $\Delta$6

$\Delta$7

$\Delta$8

PCS is then regressed from the deepest fused representation: $\Delta$9

The training objective combines sigmoid focal loss for classification, DIoU loss for localization, MSE for element-score regression, and MSE for PCS regression. The paper reports that the best balancing coefficients are $B$0 and $B$1 (Wang et al., 22 Aug 2025).

4. Variants in state-space modulation and pyramid interaction

In neighboring work, the most technically consequential differences lie in how the second stream interacts with Mamba’s state-space mechanism. HiFi-Mamba is a dual-stream MRI reconstruction network in which the incoming feature map is channel-split into two halves. One half is passed through a $B$2-Laplacian block to produce a low-frequency component and a residual high-frequency component, while the second half is added to the high-frequency branch. The high-frequency stream is refined into an anatomical guidance feature $B$3, and this guidance modulates the selective state-space branches of the low-frequency Mamba path: $B$4 The paper anchors this design in the continuous-time equations

$B$5

and reports that gating $B$6 is worse than selectively modulating $B$7 and $B$8 (Chen et al., 7 Aug 2025).

CS-MUNet adapts the same state-space formalism in two orthogonal ways. In BASM, a boundary posterior map is used to modify the spatial scan parameters at decoder skip fusion: $B$9 In CMSA, the channel dimension is redefined as the SSM sequence axis, and grouped bounded recurrence is applied over channels rather than spatial tokens. This makes the two streams dimension-oriented rather than modality-oriented: one stream models cross-channel anatomical semantic collaboration, and the other performs boundary-aware spatial fusion (Zheng et al., 20 Mar 2026).

MDN uses a less entangled dual-stream pattern. Its spatial stream is a VM-UNet/VSS-block-based U-shaped encoder-decoder, whereas its spectral stream serializes the spectral dimension and applies a Spectral Sequence Representation Layer. The core recurrence is

$T, T/2, T/4, T/8, T/16, T/32$0

The spectral features are integrated into the spatial stream by concatenation after the second encoder stage, denoted $T, T/2, T/4, T/8, T/16, T/32$1, rather than by repeated multi-level cross-stream interaction (Lin et al., 24 Feb 2025).

PyramidMamba is complementary in the opposite direction: it is strongly pyramid-centered but only weakly two-stream. Dense Spatial Pyramid Pooling constructs a large set of pooled-and-upsampled multiscale features from a single high-level feature map,

$T, T/2, T/4, T/8, T/16, T/32$2

then flattens them,

$T, T/2, T/4, T/8, T/16, T/32$3

and fuses them using a standard Mamba block,

$T, T/2, T/4, T/8, T/16, T/32$4

Its main claim is that selective scanning compresses homogeneous pyramid semantics and extracts core semantic information during multi-scale fusion (Wang et al., 2024).

5. Empirical behavior across tasks

The explicit Two-Stream Mamba Pyramid Network for figure skating reports state-of-the-art performance on FineFS. On free skating it achieves $T, T/2, T/4, T/8, T/16, T/32$5 and $T, T/2, T/4, T/8, T/16, T/32$6; on short program it achieves $T, T/2, T/4, T/8, T/16, T/32$7 and $T, T/2, T/4, T/8, T/16, T/32$8. Transfer without retraining gives TES/PCS scores of $T, T/2, T/4, T/8, T/16, T/32$9 on Fis-V and $P_i = (s_i, e_i, c_i, r_i),$0 on FS1000. Its ablations are closely aligned with the architectural thesis: w/o Audio gives TES $P_i = (s_i, e_i, c_i, r_i),$1, PCS $P_i = (s_i, e_i, c_i, r_i),$2; Symmetrical Fusion gives TES $P_i = (s_i, e_i, c_i, r_i),$3, PCS $P_i = (s_i, e_i, c_i, r_i),$4; One Stream Fusion gives TES $P_i = (s_i, e_i, c_i, r_i),$5, PCS $P_i = (s_i, e_i, c_i, r_i),$6; and Two Stream Fusion gives TES $P_i = (s_i, e_i, c_i, r_i),$7, PCS $P_i = (s_i, e_i, c_i, r_i),$8. Fusion at Levels 1–6 yields the best PCS Spearman $P_i = (s_i, e_i, c_i, r_i),$9, and using video as query with audio as key/value is better than the reverse. The Mamba temporal encoder also clearly exceeds convolutional baselines: Mamba(512) yields TES $s_i$0, PCS $s_i$1, and mAP/tIoU $s_i$2 at $s_i$3, compared with TES/PCS $s_i$4 and $s_i$5 for the two convolutional baselines (Wang et al., 22 Aug 2025).

HiFi-Mamba provides evidence for the two-stream but only weakly pyramid-like interpretation in MRI reconstruction. HiFi-Mamba(P1) achieves fastMRI results of $s_i$6 PSNR, $s_i$7 SSIM, and $s_i$8 NMSE at AF $s_i$9; on CC359 it achieves $e_i$0 PSNR, $e_i$1 SSIM, and $e_i$2 NMSE at AF $e_i$3. Against the four-directional Mamba baseline LMO on fastMRI AF $e_i$4, HiFi-Mamba(P1) uses $e_i$5G FLOPs and reaches $e_i$6 PSNR, $e_i$7 SSIM, compared with LMO’s $e_i$8G FLOPs, $e_i$9 PSNR, and $c_i$0 SSIM. The component ablation on CC359 AF $c_i$1 also shows a cumulative gain from $c_i$2-Laplacian only, to $c_i$3 HiFi-Mamba, to $c_i$4 DSFA, to $c_i$5 CRM (Chen et al., 7 Aug 2025).

MDN demonstrates that dual-stream Mamba designs also extend to hyperspectral segmentation. The model reaches DSC $c_i$6, HD $c_i$7, Throughput $c_i$8 images/s, and MACs $c_i$9G on MDC, and DSC $r_i$0, HD $r_i$1, Throughput $r_i$2 images/s, and MACs $r_i$3G on HCC. Its fusion-location ablation shows that inserting the spectral stream at $r_i$4 only gives DSC $r_i$5, outperforming $r_i$6, $r_i$7, $r_i$8, and all-layer insertion. Its spectral-module ablation is especially relevant: plain Mamba yields DSC $r_i$9, whereas SSRL yields $C$00, which the authors interpret as evidence that vanilla Mamba’s unidirectional aggregation is suboptimal for spectral-band context modeling (Lin et al., 24 Feb 2025).

In remote sensing segmentation, PyramidMamba shows that pyramid construction and Mamba fusion can be highly effective even without a fully explicit two-stream topology. It reports $C$01 mIoU on OpenEarthMap, $C$02 mIoU on Vaihingen, and $C$03 mIoU on Potsdam. On Vaihingen, the ablation Baseline $C$04 Baseline + DSPP $C$05 Baseline + DSPP + PFM improves mIoU from $C$06 to $C$07 to $C$08, and adding low-level detailed feature fusion increases mIoU further from $C$09 to $C$10 (Wang et al., 2024).

6. Conceptual boundaries, misconceptions, and limitations

A common misconception is to treat “two-stream,” “Mamba,” and “pyramid” as if they automatically imply one fixed network topology. The literature does not support that reading. Two-stream may refer to modality separation, as in video versus audio-visual figure-skating assessment (Wang et al., 22 Aug 2025); frequency separation, as in low-/high-frequency MRI reconstruction (Chen et al., 7 Aug 2025); spatial versus spectral processing, as in hyperspectral segmentation (Lin et al., 24 Feb 2025); or channel versus spatial state-space modeling, as in abdominal segmentation (Zheng et al., 20 Mar 2026). Likewise, “pyramid” may refer to a literal multilevel temporal hierarchy, a dense pooled multiscale decoder, a U-shaped multiresolution scaffold, or only a loose analogy to Laplacian residual decomposition.

Another misconception is to assume that any paper mentioning “multi-scale fusion” has specified a formal pyramid network. UltrasODM is illustrative: it states that the method “concludes with a multi-scale fusion strategy for 6-DoF parameter regression,” but the paper does not define explicit pyramid levels, top-down lateral fusion, or a formal feature pyramid architecture. It is best read as support for dual-stream Mamba temporal fusion rather than as a validated pyramid-network formulation (Anand et al., 8 Dec 2025).

Several papers also leave important implementation details under-specified. HiFi-Mamba does not provide the full selective scan recurrence or an explicit data-consistency formula, even though data consistency is central to the reconstruction loop (Chen et al., 7 Aug 2025). The figure-skating paper uses Mamba throughout the hierarchy but does not derive task-specific state-space equations beyond module placement and ablation evidence (Wang et al., 22 Aug 2025). PyramidMamba makes the flatten-then-fuse logic explicit, but the exact four-route expansion and merge details are not fully formalized in the extracted text (Wang et al., 2024). This suggests that the term remains more descriptive than standardized: it identifies a recurrent design tendency—dual complementary streams plus Mamba-based long-range modeling plus hierarchical multi-scale processing—rather than a settled canonical blueprint.