Saliency-Guided Mamba Block (SGMB) Overview

• The paper introduces SGMB as a saliency-guided state-space module that modulates recurrence via spatial and temporal cues, boosting task performance across domains.
• SGMBs employ methods such as spatial reordering, keyframe weighting, and spiking masks to enhance foreground coherence and prevent historical forgetting.
• Empirical studies demonstrate SGMB's efficiency, showing improved metrics like higher structure-measure and lower FID compared to standard models and Transformers.

A Saliency-Guided Mamba Block (SGMB) is a state-space neural network module that integrates task-driven saliency signals into the internal sequential processing of the Mamba architecture. SGMBs are designed to enhance a Mamba block’s ability to selectively emphasize foreground regions or key temporal moments, improving both spatial/temporal coherence and task performance in settings such as salient object detection, video grounding, and text-to-motion generation. The SGMB concept has been instantiated in varied domains, notably in the Samba+ salient object detection framework (Zhao et al., 2 Feb 2026), the T2M Mamba for text-driven human motion generation (Zhan et al., 1 Feb 2026), and temporal video grounding with SpikeMba (Li et al., 2024).

1. Architectural Principles and Rationale

The primary innovation of the SGMB lies in its direct use of saliency cues to modulate the recurrence mechanism of the state-space model (SSM). Unlike conventional Mamba or Visual State-Space (VSS) blocks—where sequential updates operate over uniformly ordered patches or tokens—SGMBs introduce algorithmic modifications so that state transitions preferentially process salient spatial regions or temporally significant frames. This guidance can be realized through spatial reordering (Samba+), temporal spiking or masking (SpikeMba), or explicit weighting and phase-biasing (T2M Mamba).

This design enables SGMBs to:

• Preserve foreground coherence by aligning state updates along the geometric path of salient objects.
• Reinforce representation of salient or “key” frames in long sequences, preventing historical forgetting.
• Inject inductive biases for periodicity and other task-specific structural priors.

2. Core SGMB Variants Across Application Domains

Three distinct SGMB instantiations have been documented, each tailored to its domain:

Application	Saliency Mechanism	Key Modulation in Mamba Block
Salient Object Detection	Spatial Neighborhood Scanning (SNS)	Saliency-driven sequence ordering
Text-to-Motion Generation	Keyframe weights + Fourier phase encodings	Weighted input/projected phases
Temporal Video Grounding	Binary spiking saliency mask (LIF neuron)	Explicit mask gating, slot fusion

Samba+/SGMB: Salient object detection employs an SNS algorithm to flatten spatially contiguous salient patches into 1D processing paths. Four path variants (forward, backward, mixed) are constructed using thresholded coarse saliency maps (Zhao et al., 2 Feb 2026).

T2M Mamba SGMB: For motion generation, enhanced Density Peaks Clustering estimates keyframe saliency weights, while periodicity is injected via Fourier phase encodings. The recurrence is modulated by both saliency and phase-derived embeddings (Zhan et al., 1 Feb 2026).

SpikeMba SGMB: In temporal video grounding, a spiking saliency detector produces binary proposal sets via Leaky Integrate-and-Fire neurons. “Relevant slots” maintain context, and saliency masks modulate state-space updates with language guidance (Li et al., 2024).

3. Mathematical Description

Samba+ Saliency-Guided Spatial Ordering

Given a coarse saliency map $S_c \in \mathbb{R}^{H \times W}$:

• Compute binary mask $M$ (threshold at $0.5$).
• Define salient patch set $P = \{(i,j)\mid M_{i,j}=1\}$.
• Construct path $I$ via row-wise, alternating left–right or right–left scans, ensuring spatial contiguity between consecutive salient patches:

$$\text{minimize} \sum_{k=1}^{|P|-1} \|\pi(k+1) - \pi(k)\|_2$$

• Generate four scan variants. Each 1D sequence is processed by an S6 selective state-space block, outputs are remapped to 2D and summed (Zhao et al., 2 Feb 2026).

T2M Mamba Keyframe/Periodicity Coupling

• Keyframe saliency $F \in \mathbb{R}^{L \times 1}$ via Density Peaks Clustering.
• Periodicity $\Phi \in \mathbb{R}^{L \times 2}$ via FFT autocorrelation, forming phase codes $[\sin(2\pi t/T), \cos(2\pi t/T)]$.
• Input enhancement:

$X_\Phi = X + \Phi W_\phi$

• Saliency-gated input projection:

$\overline{B}_k = \overline{B} \odot F$

• SGMB recurrence:

$$h_t = \overline{A}\, h_{t-1} + \overline{B}_k\, [X_\Phi]_t, \quad y_t = C h_t$$

[Pseudocode given in the original, see (Zhan et al., 1 Feb 2026).]

SpikeMba Temporal Saliency Masking

• Saliency mask $S[t]$ via spiking neuron:

$$U[t]=H[t-1]+X[t], \quad S[t]=\mathrm{Hea}(U[t]-U_{\mathrm{th}})$$

• Mask interacts with contextual slots and SSM layers, forming a fused update modulated by language gating (Li et al., 2024).

4. Algorithmic Pipeline and Pseudocode

A canonical SGMB in the spatial SOD context (Samba+) includes normalization, two-branch feature processing, saliency-guided spatial order computation (SNS), SSM recurrence along guided sequences, and branchwise merging:

Input: Feature F [H, W, C], saliency map S_c [H, W] 1. U = LayerNorm(F) 2. B1 = Linear(U); B1 = DWConv(B1); B1 = SiLU(B1) 3. M = (S_c > 0.5); I = SNS(M) 4. for p in 1..4: x_p = flatten F along I^(p) y_p = S6(x_p) Y_p = map back to [H, W, C] 5. X_out = sum_p Y_p 6. T = LayerNorm(X_out) 7. F' = F + Linear(T) ⊙ B2 # B2: secondary conv branch Return F'

Analogous pipelines hold in the temporal and multi-modal variants, with domain-specific saliency computation and gating.

5. Computational Complexity and Efficiency

SGMBs preserve the linear complexity $O(NC)$ (with $N$ the number of patches/frames and $C$ the channel dimension) of vanilla Mamba/VSS blocks. The SNS reordering is $O(N)$ and the four-path recursion introduces a small constant factor ($\approx 8NC$ vs $4NC$ in the standard path-agnostic case). Compared to Transformer-based self-attention, whose complexity is $O(N^2C)$, SGMBs are substantially more efficient in hardware and wall-clock terms (Zhao et al., 2 Feb 2026).

The additional overhead for saliency computation (SNS, DPC, LIF neuron masking) is negligible at typical image or sequence sizes. Empirical results confirm that the architectural overhead is ≤20% above a vanilla state-space block, with absolute runtimes still well below N² attention models.

6. Empirical Validation and Ablation Studies

Ablations demonstrate the necessity and effect of the saliency-guided mechanism:

• In Samba+, replacing SGMB with standard scan directions or removing it lowers structure-measure $(S_m)$ by 1–3 points across six SOD datasets (e.g., DUTS: $S_m$ from $0.936$ to $0.922$ with SGMB removed) (Zhao et al., 2 Feb 2026).
• In T2M Mamba, FID (lower is better) on HumanML3D drops from $0.125$ (no saliency/phase) to $0.068$ (with SGMB); ablating saliency weighting or phase injection individually increases FID by $29\%$ and $16\%$, respectively (Zhan et al., 1 Feb 2026).
• For SpikeMba, ablating the spiking saliency detector drops mean Average Precision by 3.3 points, and removing slot-guided SSM further erodes performance (Li et al., 2024).

7. Broader Implications and Limitations

Saliency-Guided Mamba Blocks provide a principled mechanism for injecting both top-down (task-driven) and bottom-up (data-driven) saliency into sequence-processing neural models, while retaining the scalability and global context benefits of state-space architectures. The adoption of SGMBs in multi-modal, vision, and generative tasks underlines their task versatility.

Current limitations include the heuristic nature of certain saliency selection procedures (threshold choices in SNS, DPC peak selection), the manual tuning of spiking/slot parameters, and the integration challenge between heterogeneous modules (SNNs vs SSMs). Further research into differentiable saliency path selection and joint SSM-SNN architectures may further broaden the applicability and robustness of SGMBs.

---

For comprehensive architectural, mathematical, and empirical details, see the corresponding sections, ablation tables, and pseudocode in (Zhao et al., 2 Feb 2026, Zhan et al., 1 Feb 2026), and (Li et al., 2024).