Vcamba: Efficient Camouflaged Video Detection

• Vcamba is a video camouflaged object detection framework that uniquely combines state-space modeling with dual-domain motion analysis for enhanced motion perception.
• The architecture employs a U-shaped encoder-decoder with bifurcated spatial and frequency branches, enabling efficient fusion of multi-scale features and reduced computational cost.
• It achieves superior performance in metrics such as Sα and Fβ^w while halving MACs compared to previous methods, making it a practical solution for video camouflage detection.

Searching arXiv for the specified paper and closely related Mamba-based vision/camouflage works to ground the article in current literature. Visual Camouflage Mamba (Vcamba) is a video camouflaged object detection (VCOD) architecture that combines state space modeling with dual-domain motion analysis in spatial and frequency representations. It is presented as a Mamba-based efficient spatio-frequency motion perception framework for breaking camouflage in video, where foreground and background exhibit high similarity in color and texture, and it integrates frequency and spatial features for efficient and accurate VCOD (Li et al., 31 Jul 2025).

1. Research context and nomenclature

Vcamba belongs to a broader line of work that applies State Space Models (SSMs), especially Mamba-style selective scanning, to vision tasks in which long-range dependencies and sequence efficiency are central. In the formulation associated with Vcamba, the key motivation is that existing VCOD methods primarily rely on spatial appearance features to perceive motion cues, but the high similarity between foreground and background in VCOD results in limited discriminability of spatial appearance features, restricting detection accuracy and completeness. The design therefore introduces frequency features not only to enhance feature representation but also to perceive motion through dynamic variations in frequency energy (Li et al., 31 Jul 2025).

The name is easily confused with other Mamba-based visual systems. “VCMamba” denotes a distinct vision backbone that bridges convolutions with multi-directional Mamba for image classification and semantic segmentation rather than video camouflaged object detection (Munir et al., 4 Sep 2025). “Mamba Capsule Routing Towards Part-Whole Relational Camouflaged Object Detection” addresses image-level camouflaged object detection through type-level capsule routing and Capsules Spatial Details Retrieval, not video motion modeling (Zhang et al., 2024). A common misconception is therefore to treat Vcamba as a generic visual backbone or as a capsule-routing method; in the cited literature, it is specifically a U-shaped encoder–decoder VCOD model with a dual-branch motion perception neck (Li et al., 31 Jul 2025).

This placement is technically significant because the three systems emphasize different inductive biases. VCMamba emphasizes hybrid local/global representation in a hierarchical vision backbone (Munir et al., 4 Sep 2025). Mamba Capsule Routing emphasizes part-whole relational reasoning at the type level (Zhang et al., 2024). Vcamba emphasizes efficient motion perception across frame sequences in both spatial and frequency phase domains (Li et al., 31 Jul 2025). This suggests that the shared use of Mamba does not imply architectural equivalence; rather, the SSM formalism is adapted to distinct sequence constructions and task priors.

2. System architecture and data flow

Vcamba adopts a U-shaped encoder–decoder structure with a dual-branch “motion perception” neck. Given a video clip of $N$ frames $\{I_t\}_{t=1}^N$, each $3 \times H \times W$, the model begins with an ImageNet-pretrained Vision Mamba (VMamba) backbone with four hierarchical Visual State Space (VSS) layers. In each VSS layer, the standard feed-forward network is replaced by a Receptive Field VSS (RFVSS) block that injects multi-scale convolutional receptive fields into the SSM-based state updater. The encoder outputs four multi-scale feature tensors $$\{F_i \in \mathbb{R}^{N \times C_i \times H_i \times W_i}\}_{i=1}^4$$ (Li et al., 31 Jul 2025).

The motion perception neck is bifurcated into a spatial branch and a frequency branch. In the spatial branch, Difference-based Space Enhancement (DSE) computes per-frame inter-feature differences $F_i^D = F_{i+1} - F_i$, then self-attends to yield enhanced spatial cues $F_i^{SE}$. These features are processed by the Space-based Long-Range Motion Perception (SLMP) module, which serializes the spatio-temporal map into 1D sequences via four directional scans—horizontal forward, horizontal backward, vertical forward, and vertical backward. Each serialized sequence is processed by an SSM block and then re-merged into per-frame spatial motion features $F_i^{spa} \in \mathbb{R}^{C \times H \times W}$ (Li et al., 31 Jul 2025).

The frequency branch begins with Adaptive Frequency Component Enhancement (AFE). AFE transforms each $F_i$ via channel-wise 2D-FFT into amplitude and phase maps, then “spiral-scans” the 2D frequency map low$\to$high and high$\to$low into two 1D sequences. These sequences are modeled by an SSM to adaptively weight and enhance frequency bands, then re-assembled and iFFT-reconstructed to produce enhanced frequency features $\{I_t\}_{t=1}^N$0. Frequency-based Long-Range Motion Perception (FLMP) then takes the phase-only maps of $\{I_t\}_{t=1}^N$1, stacks them into a $\{I_t\}_{t=1}^N$2 spatio-frequency tensor, runs four-directional SSM scans to yield per-frame phase-motion features $\{I_t\}_{t=1}^N$3, and uses a small Mamba-fusion module (MFM) to gate amplitude maps $\{I_t\}_{t=1}^N$4 via a sigmoid, producing guided amplitude $\{I_t\}_{t=1}^N$5. Finally, $\{I_t\}_{t=1}^N$6 yields per-frame frequency-motion features $\{I_t\}_{t=1}^N$7 (Li et al., 31 Jul 2025).

The two branches remain separate until the Space–Frequency Motion Fusion (SFMF) stage. There, $\{I_t\}_{t=1}^N$8 and $\{I_t\}_{t=1}^N$9 are flattened and concatenated in two ways: sequence-to-sequence for global interaction and point-to-point along patch correspondences for local interaction. The concatenated 1D sequences pass through a shared SSM scan, cross-merge, a linear layer with FFN, and a final $3 \times H \times W$0 convolution to yield a unified motion representation $3 \times H \times W$1 for each frame. Four VSS decoder layers then mirror the encoder, up-sampling and fusing across scales to predict per-frame camo-object masks $3 \times H \times W$2 (Li et al., 31 Jul 2025).

3. State-space formulation and module-level mathematics

The SSM foundation of Vcamba is given in continuous time by

$3 \times H \times W$3

where $3 \times H \times W$4 is the hidden state, $3 \times H \times W$5 the input signal, $3 \times H \times W$6 the output, and $3 \times H \times W$7 are the state-space parameters. Discretizing via zero-order hold with step $3 \times H \times W$8 yields

$3 \times H \times W$9

followed by the recursion

$$\{F_i \in \mathbb{R}^{N \times C_i \times H_i \times W_i}\}_{i=1}^4$$0

In an S6 (Selective) block, $$\{F_i \in \mathbb{R}^{N \times C_i \times H_i \times W_i}\}_{i=1}^4$$1, $$\{F_i \in \mathbb{R}^{N \times C_i \times H_i \times W_i}\}_{i=1}^4$$2, and $$\{F_i \in \mathbb{R}^{N \times C_i \times H_i \times W_i}\}_{i=1}^4$$3 become input-dependent parameters generated via small neural nets of $$\{F_i \in \mathbb{R}^{N \times C_i \times H_i \times W_i}\}_{i=1}^4$$4, allowing dynamic focus (Li et al., 31 Jul 2025).

Within RFVSS, if $$\{F_i \in \mathbb{R}^{N \times C_i \times H_i \times W_i}\}_{i=1}^4$$5 is the input at layer $$\{F_i \in \mathbb{R}^{N \times C_i \times H_i \times W_i}\}_{i=1}^4$$6, time $$\{F_i \in \mathbb{R}^{N \times C_i \times H_i \times W_i}\}_{i=1}^4$$7, the module first performs channel expansion and activation:

$$\{F_i \in \mathbb{R}^{N \times C_i \times H_i \times W_i}\}_{i=1}^4$$8

It then applies multi-scale depthwise-separable convolutions in parallel,

$$\{F_i \in \mathbb{R}^{N \times C_i \times H_i \times W_i}\}_{i=1}^4$$9

and aggregates them as

$F_i^D = F_{i+1} - F_i$0

The state update becomes

$F_i^D = F_{i+1} - F_i$1

This module therefore injects multi-scale convolutional receptive fields into the SSM-based state updater rather than using a standard FFN (Li et al., 31 Jul 2025).

AFE operates in the frequency domain. For a spatial feature $F_i^D = F_{i+1} - F_i$2, channel-wise FFT is defined as

$F_i^D = F_{i+1} - F_i$3

The result is separated into magnitude $F_i^D = F_{i+1} - F_i$4 and phase $F_i^D = F_{i+1} - F_i$5. The 2D frequency map is “spiral-scanned” into a 1D sequence $F_i^D = F_{i+1} - F_i$6 in low$F_i^D = F_{i+1} - F_i$7high order, and also in reverse high$F_i^D = F_{i+1} - F_i$8low order. Each sequence is passed through an SSM:

$F_i^D = F_{i+1} - F_i$9

The outputs are scattered back into a $F_i^{SE}$0 frequency map $F_i^{SE}$1, and the enhanced feature is reconstructed as

$F_i^{SE}$2

The stated purpose is to adaptively weight and enhance frequency bands while maintaining semantic consistency through the scanning strategy (Li et al., 31 Jul 2025).

SLMP and FLMP instantiate long-range motion modeling in the spatial and frequency phase domains, respectively. In SLMP, each spatial feature map is flattened and stacked into a tensor of shape $F_i^{SE}$3, after which four directional SSM scans are applied over the serialized $F_i^{SE}$4 tokens:

$F_i^{SE}$5

Each output sequence is reshaped to $F_i^{SE}$6, fused, and unstacked to per-frame motion features. In FLMP, only phases are flattened and stacked; the resulting phase-motion features $F_i^{SE}$7 guide amplitudes through

$F_i^{SE}$8

and the reconstructed frequency-motion feature is

$F_i^{SE}$9

SFMF then fuses $F_i^{spa} \in \mathbb{R}^{C \times H \times W}$0 and $F_i^{spa} \in \mathbb{R}^{C \times H \times W}$1 through both sequence-to-sequence concatenation $F_i^{spa} \in \mathbb{R}^{C \times H \times W}$2 and point-to-point concatenation over patch correspondences, followed by SSM scanning, a linear layer, FFN, and a final $F_i^{spa} \in \mathbb{R}^{C \times H \times W}$3 convolution to obtain the fused map $F_i^{spa} \in \mathbb{R}^{C \times H \times W}$4 (Li et al., 31 Jul 2025).

4. Supervision, optimization targets, and evaluation criteria

Vcamba is trained with a hybrid loss that combines weighted BCE, weighted IoU, and enhanced-alignment loss:

$F_i^{spa} \in \mathbb{R}^{C \times H \times W}$5

Across four scales and $F_i^{spa} \in \mathbb{R}^{C \times H \times W}$6 frames, the total objective is

$F_i^{spa} \in \mathbb{R}^{C \times H \times W}$7

The formulation is explicitly per frame and per scale, which aligns the optimization target with the multiscale decoder outputs (Li et al., 31 Jul 2025).

The reported evaluation uses six standard metrics for a predicted mask $F_i^{spa} \in \mathbb{R}^{C \times H \times W}$8 and ground-truth $F_i^{spa} \in \mathbb{R}^{C \times H \times W}$9. These are structure measure $F_i$0, weighted F-measure $F_i$1, enhanced-alignment $F_i$2, mean absolute error $F_i$3, meanDice $F_i$4, and meanIoU $F_i$5. The metric definitions are given explicitly:

$F_i$6

$F_i$7

$F_i$8

$F_i$9

$\to$0

These choices reflect an emphasis on both segmentation fidelity and structural completeness, which is especially relevant in camouflage settings where boundary ambiguity and incomplete object recovery are recurrent failure modes (Li et al., 31 Jul 2025).

5. Reported performance and computational profile

The quantitative results reported for Vcamba on the MoCA-MASK test set are as follows.

Metric / cost	Vcamba
$\to$1	0.684
$\to$2	0.382
$\to$3	0.804
$\to$4	0.010
$\to$5	0.459
$\to$6	0.369
MACs on $\to$7 input	10.88 G

Against the prior SLTNet on MoCA-MASK, the reported gains are $\to$8, $\to$9, $\to$0, $\to$1, $\to$2, and $\to$3. The paper further states that similar large gains appear on CAD2016. It also reports that Vcamba halves MACs compared to SLTNet, with $\to$4 versus $\to$5, and compares against EMIP at $\to$6 (Li et al., 31 Jul 2025).

The complexity claims are tied directly to the scan-based SSM design. SSM blocks in Mamba scan a 1D sequence of length $\to$7 in $\to$8 time and $\to$9 memory, plus hidden state size $\{I_t\}_{t=1}^N$00, whereas a standard Transformer self-attention on length $\{I_t\}_{t=1}^N$01 costs $\{I_t\}_{t=1}^N$02. For a feature map of $\{I_t\}_{t=1}^N$03 over $\{I_t\}_{t=1}^N$04 frames, each directional Mamba scan costs $\{I_t\}_{t=1}^N$05, and there are a fixed constant number of scans—stated as $\{I_t\}_{t=1}^N$06 across SLMP, FLMP, SFMF, AFE, and RFVSS combined—so the total time remains $\{I_t\}_{t=1}^N$07 (Li et al., 31 Jul 2025).

A plausible implication is that the efficiency gains are not derived from removing temporal modeling, but from replacing heavier pairwise or kernel-based sequence operators with a constant-number scan regime over serialized spatio-temporal and frequency-temporal sequences. In the reported framing, this is the basis for combining long-range motion perception with lower computation cost.

6. Technical significance and relation to adjacent methods

The defining technical feature of Vcamba is the explicit division of motion perception into a spatial branch and a frequency branch, followed by SSM-based fusion. The spatial branch models inter-frame feature differences and four-directional spatio-temporal scans, while the frequency branch models amplitude and phase behavior after channel-wise 2D-FFT, including spiral frequency-domain sequential scanning and phase-guided amplitude modulation (Li et al., 31 Jul 2025). This suggests that Vcamba treats camouflage breaking not as a purely appearance-based segmentation problem, but as a joint motion-and-spectrum inference problem in which phase-domain dynamics can supply cues that spatial appearance alone cannot provide.

This orientation distinguishes Vcamba from image-only camouflage systems. Mamba Capsule Routing uses a Swin-Transformer encoder, Mamba capsule generation, type-level EM routing, and Capsules Spatial Details Retrieval for camouflaged object detection on still images, with datasets such as CAMO, COD10K, and NC4K (Zhang et al., 2024). Vcamba, by contrast, is organized around per-frame mask prediction in video, dual-domain motion perception, and a hybrid multiscale supervision strategy (Li et al., 31 Jul 2025). The two approaches therefore address related camouflage phenomena but operate on different task formulations and different abstractions of long-range dependency.

Vcamba is also distinct from general-purpose Mamba vision backbones. VCMamba integrates a convolutional stem, hierarchical convolutional FFN stages, and multi-directional Mamba blocks for ImageNet-1K classification and ADE20K semantic segmentation, emphasizing a hybrid local/global representation with linear complexity in image resolution (Munir et al., 4 Sep 2025). Vcamba instead uses VMamba encoder layers modified by RFVSS, then introduces specialized SLMP, FLMP, and SFMF modules tailored to video motion cues under camouflage (Li et al., 31 Jul 2025). A common misreading is therefore to collapse all “visual Mamba” models into a single architectural family; the cited literature indicates that the shared SSM machinery supports sharply different task-specific sequence constructions, from image token scans to type-level capsule routing to spatio-frequency video motion modeling.

In that sense, Vcamba occupies a specific niche within Mamba-based vision research: it is neither a generic backbone nor an image-only camouflage detector, but a task-specific VCOD architecture whose central claim is that efficient and accurate camouflaged video segmentation can be achieved by coupling multi-scale receptive-field VSS blocks with adaptive frequency-domain enhancement, dual-domain long-range motion perception, and SSM-based fusion (Li et al., 31 Jul 2025).