Hierarchical Alternate Interaction Network for RGB-D Salient Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing RGB-D Salient Object Detection (SOD) methods often overlook depth map quality.
• Depth maps can contain distractors due to acquisition issues, impacting detection accuracy.

Purpose of the Study:

• To propose a novel Hierarchical Alternate Interactions Network (HAINet) for robust RGB-D SOD.
• To mitigate depth map distractors and enhance salient object detection in RGB images.

Main Methods:

• Developed HAINet with three stages: feature encoding, cross-modal alternate interaction, and saliency reasoning.
• Introduced the Hierarchical Alternate Interaction Module (HAIM) for progressive RGB-depth-RGB feature refinement.
• Employed a hybrid loss function for effective model training.

Main Results:

• HAINet demonstrated competitive performance against 19 state-of-the-art methods across seven datasets.
• Achieved a real-time processing speed of 43 frames per second on a single NVIDIA Titan X GPU.
• Successfully filtered distractors in depth maps, improving salient object detection.

Conclusions:

• HAINet offers an effective solution for RGB-D Salient Object Detection by addressing depth map quality issues.
• The proposed method achieves both high accuracy and real-time processing capabilities.