MobileSal: Extremely Efficient RGB-D Salient Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• High computational cost of neural networks hinders real-world applications of RGB-D salient object detection (SOD).
• Existing methods often rely on cumbersome networks, limiting their practical deployment.

Purpose of the Study:

• To introduce MobileSal, an efficient network for RGB-D SOD designed for mobile devices.
• To enhance feature representation capabilities of mobile networks for improved SOD performance.
• To enable real-time and resource-efficient salient object detection.

Main Methods:

• Utilized mobile networks for deep feature extraction in RGB-D SOD.
• Proposed an Implicit Depth Restoration (IDR) technique to strengthen feature representation during training (computationally free during testing).
• Developed Compact Pyramid Refinement (CPR) for efficient multi-level feature aggregation and precise boundary delineation.

Main Results:

• MobileSal demonstrates favorable performance against state-of-the-art methods on six challenging RGB-D SOD datasets.
• Achieved significantly faster speed (450fps at 320x320 resolution) and fewer parameters (6.5M).
• Successfully enabled efficient and accurate salient object detection.

Conclusions:

• MobileSal offers a computationally efficient and effective solution for RGB-D SOD on mobile platforms.
• The integration of IDR and CPR significantly boosts the performance of mobile networks for this task.
• The proposed method paves the way for practical, real-world applications of salient object detection.