Learning global-view correlation for salient object detection in 3D point clouds.

Area of Science:

• Computer Vision
• 3D Data Analysis
• Artificial Intelligence

Background:

• Salient object detection (SOD) in 3D point clouds is challenging due to data irregularity.
• Existing methods often miss crucial scene-level global-view correlations.
• Effective saliency prediction requires understanding both local and global context.

Purpose of the Study:

• To develop a novel approach for salient object detection in 3D point clouds.
• To address the limitations of existing methods by incorporating global-view scene understanding.
• To introduce a method that refines saliency representations by filtering scene-common correlations.

Main Methods:

• Proposed the Saliency Filtration Network (SFN) with a two-stage strategy.
• Introduced the Residual Relation-aware Transformer (RRT) module for long-range context aggregation.
• Developed the Global Bilinear Correlation based Filtration (GBCF) module for saliency purification.

Main Results:

• SFN effectively isolates salient objects by refining representations from global scene correlations.
• The RRT module captures long-range dependencies inspired by human visual perception.
• The GBCF module establishes dense correlations for accurate saliency purification.

Conclusions:

• The proposed SFN method achieves state-of-the-art accuracy in 3D salient object detection.
• SFN significantly outperforms existing methods on the PCSOD benchmark.
• This work advances SOD in point clouds by leveraging global-view scene understanding.