Improving RGB-D Salient Object Detection via Modality-Aware Decoder.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Existing RGB-Depth salient object detection (SOD) methods rely on cross-modal and cross-level fusion.
• Current fusion patterns are overly dependent on network adaptability and may not properly learn modality-level relationships.

Purpose of the Study:

• To address the limitations of existing SOD methods by developing a more modality-sensitive fusion approach.
• To introduce the Modality-aware Decoder (MaD) to improve RGB-D saliency prediction.

Main Methods:

• The proposed Modality-aware Decoder (MaD) incorporates feature embedding, modality reasoning, and feature back-projecting strategies.
• MaD transforms the standard multi-scale and multi-level decoding process into a modality-aware one.
• The method focuses on learning relationships at the modality level, considering factors like depth quality and scene complexity.

Main Results:

• MaD achieves competitive performance compared to state-of-the-art (SOTA) models.
• The model demonstrates improved salient object detection by effectively leveraging RGB and Depth information.
• Performance gains are achieved without employing complex or specialized decoder designs.

Conclusions:

• The Modality-aware Decoder (MaD) offers a novel and effective approach to RGB-Depth salient object detection.
• By explicitly considering modality-level interactions, MaD overcomes limitations of previous methods.
• The proposed method provides a strong baseline for future research in multi-modal saliency detection.