Liquid Fusion of Heterogeneous Representations Towards General Salient Object Detection

2026-06-25 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

The Liquid Fusion Network (LFNet) addresses a critical limitation in General Salient Object Detection (SOD) by harmonizing heterogeneous representations from State Space Models (SSMs) and Convolutional Neural Networks (CNNs). Current SOD methods neglect the inherent spectral biases of these network paradigms, which exhibit complementary frequency preferences. Inspired by Liquid Neural Networks (LNNs), LFNet introduces a liquid fusion mechanism to dynamically integrate features from VMamba (SSM) and ConvNeXt (CNN) backbones. It treats VMamba features as evolving states and ConvNeXt features as exogenous stimuli, employing a dynamic gating for content-aware aggregation. This state-stimulus paradigm enhances flexibility for multi-modal cues. Additionally, LFNet incorporates a Saliency-Guided Upsampling (SGU) operator, utilizing a spectral-spatial co-design to mitigate upsampling artifacts while preserving semantic information. Extensive experiments across five tasks—RGB, RGB-D, RGB-T, VSOD, and VDT—demonstrate LFNet's state-of-the-art performance, offering a superior balance between detection accuracy and model efficiency.

Key takeaway

For Computer Vision Engineers developing advanced Salient Object Detection (SOD) systems, LFNet presents a significant advancement. If your projects involve multi-modal data or require superior accuracy and efficiency, you should evaluate LFNet's dynamic fusion of SSM and CNN features. This approach effectively addresses spectral biases, offering state-of-the-art performance across diverse tasks like RGB-D and VSOD. Consider exploring the released code to integrate its liquid fusion and Saliency-Guided Upsampling techniques into your next-generation models.

Key insights

Harmonizing complementary spectral biases of CNNs and SSMs via dynamic liquid fusion improves salient object detection.

Principles

• Combine CNNs and SSMs for complementary spectral coverage.
• Dynamic gating can integrate heterogeneous feature types effectively.
• Spectral-spatial co-design improves upsampling quality.

Method

LFNet dynamically integrates VMamba (state) and ConvNeXt (stimulus) features using a gating mechanism, then employs Saliency-Guided Upsampling for artifact-free propagation.

In practice

• Apply LFNet for multi-modal salient object detection.
• Explore dynamic fusion for other heterogeneous model combinations.
• Utilize SGU to refine feature propagation in vision tasks.

Topics

• Salient Object Detection
• State Space Models
• Convolutional Neural Networks
• Multi-modal Fusion
• Liquid Neural Networks
• Feature Integration

Code references

• cke520/LFNet

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• From Spatial to Spectral: An Efficient, Frequency-Guided Feature Representation Learner for Small Object Detection — Shifting small object detection from spatial to spectral processing improves efficiency and accuracy by preserving high-frequency features.
• Architecture-Adaptive Uncertainty Fusion for Deepfake Detection — Correlation-Optimized Fusion (COF) adaptively combines multiple uncertainty sources to predict deepfake detection errors, improving cross-domain reliability.
• DSFNet: Learning Dual-Domain Spectral Operators for Multi-Modality Spatio-Temporal Forecasting in Urban Transportation Systems — DSFNet uses dual-domain spectral filtering to model complex spatio-temporal and cross-modality couplings in urban traffic.
• FUSE: Frequency-domain Unification and Spectral Energy Alignment for Multi-modal Object Re-Identification — FUSE improves multi-modal ReID by leveraging frequency-domain analysis for spectral disentanglement and energy alignment across modalities.
• Context-Aware Feature-Fusion for Co-occurring Object Detection in Autonomous Driving — Context-aware feature fusion using local and global attention significantly improves object detection for challenging cases in autonomous driving.
• FusionRS: A Large-Scale RGB-Infrared Remote Sensing Dataset for Dual-Modal Vision-Language Foundation Models — FusionRS enables dual-modal RGB-infrared vision-language models through a novel dataset with IR-aware captions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.