MobileSal

MobileSal performs real-time salient object detection on RGB-D images using lightweight mobile networks and techniques such as Implicit Depth Restoration (IDR) and Compact Pyramid Refinement (CPR) to improve feature representation and boundary precision.

Key Features:

• Efficient Deep Feature Extraction: Mobile networks extract deep features from RGB-D images with a lightweight architecture to reduce computational burden while maintaining robust SOD performance.
• Implicit Depth Restoration (IDR): IDR leverages depth information inherent in color images to enhance feature representation by implicitly restoring depth during training and is omitted during testing.
• Compact Pyramid Refinement (CPR): CPR aggregates and refines multi-level features across scales to produce salient objects with clear boundaries.
• Performance Metrics: Evaluated on six challenging RGB-D SOD datasets, MobileSal attains 450 frames per second for 320×320 inputs and uses 6.5 million parameters.

Scientific Applications:

• Robotics: Real-time RGB-D salient object detection to support visual perception and object-focused tasks in robotic systems.
• Augmented Reality: Fast salient region extraction from RGB-D input to assist overlay and interaction computations in AR applications.
• Autonomous Driving: Rapid detection of salient objects from RGB-D data to inform perception and decision-making in driving systems.
• Mobile and Embedded Deployment: Lightweight architecture enables deployment of RGB-D SOD on mobile devices and embedded systems with constrained compute.

Methodology:

Mobile networks perform lightweight deep feature extraction; IDR implicitly restores depth during training and is omitted at inference; CPR performs compact multi-level feature refinement to improve salient object delineation.