Lightweight RGB-D Salient Object Detection From a Speed-Accuracy Tradeoff Perspective

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current RGB-D Salient Object Detection (SOD) methods often compromise efficiency for accuracy.
• Lightweight models struggle to achieve high-precision performance in RGB-D SOD tasks.

Purpose of the Study:

• To develop a lightweight RGB-D SOD network that balances efficiency and performance.
• To address limitations in depth quality, modality fusion, and feature representation for lightweight models.

Main Methods:

• Introduced the Depth Anything Model for high-quality depth map generation, mitigating multi-modal gaps.
• Proposed a Decoupled Attention Module (DAM) for exploring intra- and inter-modal consistency.
• Developed a Dual Information Representation Module (DIRM) to enhance feature representation in lightweight backbones.
• Designed a Dual Feature Aggregation Module (DFAM) for feature aggregation in the decoder.

Main Results:

• The proposed Speed-Accuracy Tradeoff Network (SATNet) achieves state-of-the-art performance.
• SATNet demonstrates a lightweight framework with only 5.2 million parameters.
• The model achieves a high inference speed of 415 Frames Per Second (FPS).

Conclusions:

• SATNet effectively balances efficiency and accuracy in RGB-D SOD.
• The proposed modules (DAM, DIRM, DFAM) contribute to improved performance in lightweight models.
• SATNet offers a competitive solution for real-time RGB-D SOD applications.