Abstract
Existing video camouflaged object detection (VCOD) methods primarily rely on spatial appearances for motion perception. However, the high foreground-background similarity in VCOD limits the discriminability of such features (e.g. color and texture). Recent studies demonstrate that frequency features can not only compensate for appearance limitations, but also perceive motion through dynamic variations in spectral energy. Meanwhile, the emerging state space model called Mamba enables efficient motion perception in frame sequences with its linear-time long-sequence modeling capability. Motivated by this, we propose Vcamba, a visual camouflage Mamba based on spatio-frequency motion perception that integrates frequency and spatial features for efficient and accurate VCOD. Specifically, by analyzing the spatial representations of frequency components, we reveal a structural evolution pattern that emerges from the ordered superposition of components. Based on this observation, we propose a unique frequency-domain sequential scanning (FSS) strategy to unfold the spectrum. Utilizing FSS, the adaptive frequency enhancement (AFE) module employs Mamba to model the causal dependencies within sequences, enabling effective frequency learning. Furthermore, we propose a space-based long-range motion perception (SLMP) module and a frequency-based long-range motion perception (FLMP) module to model spatio-temporal and frequency-temporal sequences. Finally, the space and frequency motion fusion module (SFMF) integrates dual-domain features into unified motion representation. Experiments show that Vcamba outperforms state-of-the-art methods across 6 evaluation metrics on 2 datasets with lower computation cost, confirming its superiority.
Introduction
Detecting camouflaged objects in video sequences is a notoriously difficult task in computer vision, often described as the "Where’s Waldo" of AI. Unlike standard object detection, camouflaged targets share nearly identical textures, colors, and patterns with their backgrounds, making them nearly invisible in static images. While motion is the primary cue for breaking camouflage, existing methods face two major hurdles: the lack of discriminability in spatial appearances and the high computational complexity of modeling long-range temporal dependencies using Transformers.
To address these challenges, we introduce Vcamba (Visual Camouflage Mamba). Vcamba moves beyond simple pixel-based analysis by incorporating frequency-domain features and utilizing the efficient Mamba architecture—a linear-time State Space Model (SSM)—to process video sequences.
Methodology
Vcamba is designed as a dual-branch architecture that integrates spatial and frequency cues. The framework utilizes a Vision Mamba backbone for feature extraction, followed by specialized modules for motion perception.
Key Components:
- Space-based Long-range Motion Perception (SLMP): This module captures spatio-temporal cues across frames.
- Frequency-based Long-range Motion Perception (FLMP): This branch focuses on frequency-temporal cues. It exploits the Fourier shift theorem, which suggests that while an object's appearance (amplitude) remains stable during motion, its structural position (phase) changes linearly. This provides a cleaner signal for motion detection than raw pixels.
- Space and Frequency Motion Fusion (SFMF): This module merges the dual-domain features into a unified representation, ensuring the model understands both the object's appearance and its global structural movement.
The Frequency Advantage
A core innovation of Vcamba is the Frequency-domain Sequential Scanning (FSS) strategy. Frequency data is inherently 2D, but Mamba requires sequential input. FSS "unfolds" the spectrum by spiraling from the center (low frequencies) to the edges (high frequencies). This approach is motivated by the "structural evolution pattern" observed in the frequency domain.
As shown above, superimposing frequency components in order allows the object to emerge progressively from a coarse blur into a sharp, textured image. Mamba’s causal modeling capability is uniquely suited to capturing this evolutionary sequence, allowing for more effective frequency learning through the Adaptive Frequency Enhancement (AFE) module.
Experimental Results
Vcamba was evaluated on two primary VCOD benchmarks: MoCA-MASK and CAD2016. The model demonstrated superior performance across six evaluation metrics compared to 14 state-of-the-art methods.
- Accuracy: On the MoCA-MASK dataset, Vcamba achieved an S-measure ($S_\alpha$) of 0.684, outperforming the previous best video-based model, SLT-Net (0.656).
- Efficiency: Vcamba is significantly more efficient than its competitors. It requires only 16.49G MACs (Multiply-Accumulate operations), a fraction of the 58.18G required by SLT-Net or the 1414.12G used by TSP-SAM.
- Qualitative Performance: Visualizations indicate that Vcamba produces sharper boundaries and is less susceptible to background noise than existing spatial-only models.
Conclusion and Future Work
Vcamba represents a significant step forward in video camouflaged object detection by successfully combining frequency-domain analysis with the efficiency of Mamba. By focusing on phase variations and structural evolution patterns, the model achieves state-of-the-art accuracy with remarkably low computational overhead.
While Vcamba is mathematically efficient, current hardware-level optimizations for Mamba's selective scan kernels are still maturing. Future work will focus on improving real-world inference speeds and testing the framework's robustness against more complex dynamic camouflage scenarios, such as real-time color-changing organisms.
Summary updated