Liquid Fusion of Heterogeneous Representations Towards General Salient Object Detection

Salient Object Detection (SOD) requires capturing both broad semantic context and fine-grained structural details. While recent State Space Models (SSMs) like Mamba have shown promise in handling long-range dependencies with linear complexity, they often struggle with high-frequency spatial details. This paper investigates whether the spectral biases inherent in different neural network paradigms—specifically SSMs and CNNs—can be harmonized to create a more robust and flexible general SOD framework.

The authors perform a spectral analysis using Fast Fourier Transforms (FFT) on feature maps from VMamba and ConvNeXt, revealing that these architectures possess complementary frequency preferences. To exploit this, they introduce the Liquid Fusion Network (LFNet). LFNet uses a dual-stream encoder where VMamba provides continuous state-space representations and ConvNeXt provides grid-based spatial inductive biases. These features are integrated via a Liquid Fusion Module (LFM), which treats the SSM features as an evolving memory state and the CNN features as an exogenous stimulus, dynamically gating their interaction. Additionally, the authors propose a Saliency-Guided Upsampling (SGU) operator that uses a spectral-spatial co-design to restore resolution while suppressing artifacts.

LFNet achieves state-of-the-art performance across five distinct SOD tasks: RGB, RGB-D, RGB-T, Video SOD (VSOD), and Visible-Depth-Thermal (VDT) SOD. By dynamically balancing the memory-like properties of SSMs with the stimulus-like properties of CNNs, the model effectively reconciles the representational blind spots of each individual paradigm. The SGU module further ensures that the final saliency maps maintain sharp boundaries, providing a superior trade-off between detection accuracy and computational efficiency compared to existing Transformer-based or pure SSM-based methods.