EM-Trans: Edge-Aware Multimodal Transformer for RGB-D Salient Object Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• RGB-D salient object detection (SOD) is a growing research area.
• Transformers show promise for SOD but often neglect crucial edge information, limiting accuracy.
• Existing methods struggle to effectively integrate multimodal information for precise SOD.

Purpose of the Study:

• To propose a novel edge-aware RGB-D SOD transformer (EM-Trans) that explicitly models edge information.
• To improve the accuracy of salient object detection by addressing the limitations of existing transformer models.
• To enhance the fusion of multimodal features for more robust SOD performance.

Main Methods:

• Developed EM-Trans, an edge-aware transformer utilizing a dual-band decomposition framework.
• Employed parallel decoders for high-frequency edge and low-frequency body feature extraction.
• Introduced cross-attention complementarity and color-hint guided fusion modules for feature enrichment and enhancement.
• Utilized a deeply supervised progressive fusion module for integrating edge and body features.

Main Results:

• EM-Trans effectively models edge information, overcoming limitations of prior methods.
• The proposed model achieves state-of-the-art performance on benchmark RGB-D SOD datasets, both quantitatively and qualitatively.
• Demonstrated the model's potential for broader multimodal SOD tasks, including RGB-T SOD.

Conclusions:

• EM-Trans represents a significant advancement in RGB-D salient object detection.
• Explicitly incorporating edge information is crucial for improving SOD accuracy.
• The proposed architecture offers a promising direction for future research in multimodal SOD.