Divide-and-Conquer: Confluent Triple-Flow Network for RGB-T Salient Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• RGB-Thermal Salient Object Detection (RGB-T SOD) aims to identify prominent objects in paired visible and thermal images.
• Existing methods struggle with cross-modality disparities and robustness against defective sensor data, limiting performance in complex scenarios.
• Hierarchical human visual systems inspire new approaches for robust multi-modal feature integration.

Purpose of the Study:

• To propose ConTriNet, a robust Confluent Triple-Flow Network for RGB-T SOD.
• To enhance saliency map prediction by effectively bridging RGB and Thermal modality gaps and improving robustness.
• To introduce a new benchmark dataset (VT-IMAG) for evaluating RGB-T SOD performance in challenging real-world conditions.

Main Methods:

• ConTriNet employs a unified encoder with specialized decoders using a 'Divide-and-Conquer' strategy for modality-specific and complementary information.
• Key components include a Modality-induced Feature Modulator (MFM) for reducing inter-modality discrepancies, a Residual Atrous Spatial Pyramid Module (RASPM) for multi-scale context, and a Modality-aware Dynamic Aggregation Module (MDAM).
• A flow-cooperative fusion strategy refines saliency maps from parallel triple flows (two modality-specific, one complementary) for high-resolution output.

Main Results:

• ConTriNet demonstrates superior performance over state-of-the-art methods on public benchmarks and the new VT-IMAG dataset.
• The proposed network shows significant robustness and stability, even when presented with incomplete or defective modality data.
• Experiments confirm the effectiveness of the triple-flow architecture and its specialized modules in capturing salient object information.

Conclusions:

• ConTriNet offers a robust and effective solution for RGB-Thermal Salient Object Detection, outperforming existing approaches.
• The 'Divide-and-Conquer' strategy and novel modules successfully address cross-modality challenges and improve detection accuracy.
• The developed VT-IMAG dataset provides a valuable resource for advancing research in challenging RGB-T SOD scenarios.