Dynamic Knowledge Distillation with Noise Elimination for RGB-D Salient Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• RGB-D salient object detection (SOD) uses depth information for improved accuracy in complex scenes.
• Current methods often increase model size and computation by using independent depth feature extraction streams.

Purpose of the Study:

• To develop a lightweight and efficient RGB-D SOD method.
• To reduce computational burden and model size without sacrificing detection accuracy.

Main Methods:

• Proposed a dynamic knowledge distillation (DKD) approach with a lightweight architecture.
• Dynamically assigned distillation weights based on teacher and student performance during training.
• Investigated RGB-D early fusion strategies and introduced a noise elimination method for low-quality depth maps.

Main Results:

• Achieved competitive performance across five public datasets.
• Demonstrated a fast inference speed of 136 FPS.
• Outperformed 12 prior methods in terms of efficiency and accuracy balance.

Conclusions:

• The DKD method offers an effective solution for lightweight RGB-D SOD.
• The proposed approach balances model efficiency with high detection accuracy for practical deployment.