Siamese Network for RGB-D Salient Object Detection and Beyond.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Existing RGB-D salient object detection (SOD) models often process RGB and depth data independently, limiting performance and requiring extensive training.
• This approach can be constrained by limited training data and complex training procedures.

Purpose of the Study:

• To develop a novel architecture for RGB-D salient object detection that leverages commonalities between RGB and depth modalities.
• To improve the robustness and generalization of salient object detection models.

Main Methods:

• A Joint Learning and Densely Cooperative Fusion (JL-DCF) architecture is proposed, utilizing a Siamese network for shared feature extraction from RGB and depth inputs.
• The JL module exploits cross-modal commonality for robust saliency feature learning.
• The DCF module enhances complementary feature discovery.

Main Results:

• The JL-DCF framework demonstrates robust RGB-D saliency detection with good generalization across 7 challenging datasets.
• Performance improvements of approximately 2.0% (F-measure) were achieved compared to state-of-the-art methods.
• The model showed applicability to other multi-modal detection tasks like RGB-T SOD and video SOD.

Conclusions:

• The proposed JL-DCF architecture offers significant advancements in RGB-D salient object detection.
• The JL-DCF model provides a more effective approach to multi-modal feature fusion for improved detection accuracy.
• The framework's versatility extends to related multi-modal detection tasks, demonstrating its broad utility.