MambaNeXt-YOLO: Hybrid Real-Time Detection

• The paper presents a hybrid detection architecture merging convolutional feature extraction with linear state-space modeling to efficiently capture both local and global context.
• Enhanced multi-scale fusion techniques, including asymmetric convolutions and dynamic attention, boost detection accuracy for generic, underwater, and UAV applications.
• Empirical evaluations demonstrate improved AP and FPS benchmarks compared to conventional YOLO and Transformer-based models while maintaining computational efficiency.

MambaNeXt-YOLO refers to a family of real-time object detection networks that integrate linear state space models—particularly the Mamba architecture—with deep convolutional networks in order to enhance multi-scale feature aggregation and global context modeling, while maintaining high inference efficiency. These architectures are widely applied across generic, aerial, and underwater object detection, where both accuracy and computational tractability are critical. Three representative implementations, often bearing task- or dataset-specific names, include MambaNeXt-YOLO for edge-efficient detection (Lei et al., 4 Jun 2025), SPMamba-YOLO for underwater detection (Liao et al., 26 Feb 2026), and UAVD-Mamba for multimodal UAV detection (Li et al., 1 Jul 2025).

1. Hybrid Neural-State Space Block Design

Central to MambaNeXt-YOLO architectures is the MambaNeXt block, a modular hybrid of convolutional feature extraction and Mamba-style linear state space modeling. Given an input tensor $\mathbf{X}\in\mathbb{R}^{H\times W\times C}$, the module comprises:

• Initial 1×1 projection and activation: $$\mathbf{X}' = \mathrm{SiLU}\bigl(\mathrm{BN}(\mathrm{Conv}_{1\times1}(\mathbf{X}))\bigr)$$.
• ConvNeXt local branch: Depthwise convolution, batch norm, and a two-layer MLP extract fine-grained local spatial information.
• Mamba global branch: Layer normalization and a pointwise projection yield sequences $\{\mathbf{x}_t\}_{t=1}^L$, where $L=H\times W$. Linear projections compute dynamic parameters $A_t$, $B_t$, $\Delta_t$, and hidden states update as:

$$\mathbf{h}_{t+1} = e^{-\boldsymbol{\Delta}_t}\odot\mathbf{h}_t + \mathbf{A}_t + \mathbf{B}_t\odot\mathbf{h}_t$$

The output is reshaped to a feature map and further normalized/projection-ed.

• ResGate fusion branch: Residual addition, nonlinear gating with depthwise convolutions, and additional MLPs yield gated fusions of local and global features.
• Final output: The combined output

$$\mathbf{F}_{\mathrm{final}} = \bigl(\mathbf{F}_{\mathrm{global}} + \mathbf{X}'\bigr) + \mathbf{Y}$$

ensures both spatial locality and long-range dependencies are captured efficiently.

This configuration achieves joint $\mathcal{O}(HWC)$ time complexity, enabling real-time throughput with improved semantic correspondence (Lei et al., 4 Jun 2025).

2. Multi-Scale and Asymmetric Feature Fusion Necks

MambaNeXt-YOLO designs universally employ enhanced feature pyramid networks (FPNs), such as the Multi-branch Asymmetric Fusion Pyramid Network (MAFPN) or its specialized variants. Key elements include:

• Bidirectional fusion: Top–down upsampling and bottom–up downsampling allow lateral connections between backbone features at levels $$\mathbf{X}' = \mathrm{SiLU}\bigl(\mathrm{BN}(\mathrm{Conv}_{1\times1}(\mathbf{X}))\bigr)$$0, $$\mathbf{X}' = \mathrm{SiLU}\bigl(\mathrm{BN}(\mathrm{Conv}_{1\times1}(\mathbf{X}))\bigr)$$1, $$\mathbf{X}' = \mathrm{SiLU}\bigl(\mathrm{BN}(\mathrm{Conv}_{1\times1}(\mathbf{X}))\bigr)$$2.
• Asymmetric convolutions: Replacing standard max pooling with learnable 3×3 convolutions (stride 2) for downsampling, which have been shown to improve AP$$\mathbf{X}' = \mathrm{SiLU}\bigl(\mathrm{BN}(\mathrm{Conv}_{1\times1}(\mathbf{X}))\bigr)$$3 by up to 2.5% (Lei et al., 4 Jun 2025).
• MambaNeXt block (optional) insertion: At each fusion stage, repeated application of the MambaNeXt block reinforces both local and global context propagation.
• Structural reparameterization: At inference, multi-size kernels are collapsed to single 3×3 convolutions, ensuring no additional computational overhead.

For underwater detection, the PAFPN-style neck with ODSSBlocks (containing state-space modules) is combined with spatial pyramid pooling and split-attention mechanisms (see section 3) (Liao et al., 26 Feb 2026).

3. Attention and Contextual Enhancement Mechanisms

MambaNeXt-YOLO variants augment standard object detectors with modules designed for enhanced receptive fields and feature discrimination:

• Spatial Pyramid Pooling Enhanced Layer Aggregation Network (SPPELAN): Stacks $$\mathbf{X}' = \mathrm{SiLU}\bigl(\mathrm{BN}(\mathrm{Conv}_{1\times1}(\mathbf{X}))\bigr)$$4 progressive 5×5 max-pooling layers with interleaved channel expansion and recompression ($$\mathbf{X}' = \mathrm{SiLU}\bigl(\mathrm{BN}(\mathrm{Conv}_{1\times1}(\mathbf{X}))\bigr)$$5), extending the receptive field to up to $$\mathbf{X}' = \mathrm{SiLU}\bigl(\mathrm{BN}(\mathrm{Conv}_{1\times1}(\mathbf{X}))\bigr)$$6 per dimension (Liao et al., 26 Feb 2026).
• Pyramid Split Attention (PSA): Channels are divided into $$\mathbf{X}' = \mathrm{SiLU}\bigl(\mathrm{BN}(\mathrm{Conv}_{1\times1}(\mathbf{X}))\bigr)$$7 groups, each convolved with kernels of varying sizes ($$\mathbf{X}' = \mathrm{SiLU}\bigl(\mathrm{BN}(\mathrm{Conv}_{1\times1}(\mathbf{X}))\bigr)$$8), with responses concatenated and reweighted by a global channel-wise attention vector (constructed as Squeeze-and-Excitation with sigmoid gating).
• Deformable Token Mamba Block (DTMB) and Multimodal Attentions: UAVD-Mamba replaces fixed patch tokenization with the sum of normal and deformable convolutional patches. Coupled with cross-enhanced spatial attention (using both max and mean channel pooling) and cross-channel attention (MLP-based weighting normalized across modalities), this setup facilitates superior fusion of RGB and infrared sources (Li et al., 1 Jul 2025).
• State-Space Feature Aggregation: SS2D modules implement input-adaptive, gated linear recurrences, converting standard convolutional operations into parameterized, learnable dynamical systems for spatial context propagation (Liao et al., 26 Feb 2026).

4. End-to-End Training and Loss Function Engineering

Training protocols are dataset- and model-variant dependent, but display commonalities:

• Input and augmentation: Standardized input dimensions (640×640), data augmentation (random flips, color jitter, mosaic), and normalization strategies are optimized for domain-specific robustness.
• Optimization: SGD with initial learning rates (e.g., $$\mathbf{X}' = \mathrm{SiLU}\bigl(\mathrm{BN}(\mathrm{Conv}_{1\times1}(\mathbf{X}))\bigr)$$9 or 0.01), momentum between 0.9–0.937, and weight decay $\{\mathbf{x}_t\}_{t=1}^L$0; batch sizes and epochs scaled to dataset and computational budget.
• Loss functions: Consistently, the total loss is

$\{\mathbf{x}_t\}_{t=1}^L$1

or dataset-specific extensions (e.g., inclusion of distributional Focal Loss $\{\mathbf{x}_t\}_{t=1}^L$2 for OBBs in UAVD-Mamba), with coefficients $\{\mathbf{x}_t\}_{t=1}^L$3 tuned for empirical performance.

5. Empirical Performance Across Benchmarks

The MambaNeXt-YOLO family achieves consistent improvements over both plain YOLOv8/v9/v10 backbones and contemporaneous hybrid models:

Model	Params	FLOPs	AP$\{\mathbf{x}_t\}_{t=1}^L$4 (%)	FPS (Orin/Xavier/4090)
YOLOv8-S	11.1M	28.7G	64.4	67.5 / 25.5 / 50.3
Mamba-YOLO-T	5.9M	13.6G	66.0	34.3 / 20.9 / 45.2
MambaNeXt-YOLO	7.1M	22.4G	66.6	31.9 / 19.5 / 34.6
UAVD-Mamba	39.7M	38.9G	83.0 (DV-mAP)	— / — / 14.4

For underwater detection, SPMamba-YOLO demonstrated a +4.9pt mAP@0.5 improvement over YOLOv8n (mAP 0.825 vs 0.776) with similar gains on small/dense classes (e.g., +1.5pt for Starfish), at $\{\mathbf{x}_t\}_{t=1}^L$530 FPS on an RTX 4070 (Liao et al., 26 Feb 2026). UAVD-Mamba surpassed OAFA on multimodal UAV detection by 3.6% mAP, with further ablations attributing most gain to the inclusion of DTMB and FFAR modules (Li et al., 1 Jul 2025).

Ablation studies indicate that substituting pooling with conv-based downsampling, optimizing the order and fusion of ConvNeXt and ResGate components, and scaling SSM state/width all systematically boost accuracy, with moderate impact on inference speed (Lei et al., 4 Jun 2025).

6. Domain-Specific Adaptations and Extensions

Each MambaNeXt-YOLO instantiation customizes its pipeline for its deployment scenario:

• Underwater (SPMamba-YOLO): Prioritizes perturbation-robust multi-scale enhancement (SPPELAN, PSA) and state-space modeling for resilience to low contrast and clutter (Liao et al., 26 Feb 2026).
• Generic/Edge (MambaNeXt-YOLO): Emphasizes structural efficiency, real-time throughput, and adaptability to a wide range of hardware, including NVIDIA Jetson SoCs (Lei et al., 4 Jun 2025).
• Multimodal UAV (UAVD-Mamba): Integrates deformable convolutional tokenization, cross-modal and cross-channel attention, and a YOLOv11-style neck with Mamba-augmented SPPF and C3K2 layers to maximize small object sensitivity and geometric adaptation (Li et al., 1 Jul 2025).

A plausible implication is that the MambaNeXt-YOLO paradigm can generalize across heterogeneous detection settings by parameterizing state-space modeling depth, attention mechanisms, and neck design.

7. Impact, Comparisons, and Efficiency Considerations

Hybridizing convolutional and Mamba-style state space models provides a magnet for recent research in efficient real-time detection. MambaNeXt-YOLO architectures consistently outperform both lightweight convolutional (YOLOv8n-S) and Transformer-inspired detectors, especially in settings demanding robust multi-scale or multimodal feature reasoning at real-time or near-real-time framerates (Lei et al., 4 Jun 2025, Liao et al., 26 Feb 2026, Li et al., 1 Jul 2025).

Efficiency is maintained through:

• Linear complexity state-space updates.
• Structural reparameterization post-training.
• Depthwise and asymmetric convolutions for spatial mixing.
• Channel and spatial attention for parameter-efficient context modeling.

These characteristics establish MambaNeXt-YOLO as a prototypical architecture for next-generation object detection in resource-constrained and context-sensitive environments.