Boundary-aware RGBD Salient Object Detection with Cross-modal Feature Sampling.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Mobile devices utilize depth sensors for complex tasks like salient object detection.
• Integrating data from RGB and depth sensors (RGBD) presents challenges due to differing modalities.

Purpose of the Study:

• To propose a boundary-aware cross-modal fusion network for improved RGBD salient object detection.
• To address the challenge of effectively fusing color and depth information.

Main Methods:

• Developed a cross-modal feature sampling module to balance RGB and depth feature contributions.
• Implemented a multi-scale dense fusion network incorporating edge-sensitive losses.
• Refined saliency maps using per-pixel weighted combination and an encoder-decoder network.

Main Results:

• The proposed network achieves state-of-the-art performance on public RGBD datasets.
• Demonstrated effective fusion of color and depth features for salient object detection.
• Preserved salient region boundaries through edge-sensitive losses.

Conclusions:

• The boundary-aware cross-modal fusion network significantly advances RGBD salient object detection.
• The method offers a robust solution for handling multi-modal data in mobile vision applications.