Learning of perceptual grouping for object segmentation on RGB-D data

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Object segmentation in cluttered scenes with unknown objects is a significant challenge.
• RGB-D sensors provide depth information, enabling more robust segmentation.
• Existing methods struggle with arbitrary shapes, occlusions, and complex arrangements.

Purpose of the Study:

• To introduce a novel hierarchical framework for segmenting unknown objects in RGB-D images.
• To improve object segmentation accuracy in cluttered and occluded environments.
• To leverage perceptual grouping principles for robust object hypothesis generation.

Main Methods:

• Hierarchical processing of RGB-D image data.
• Pixel-level pre-clustering and parametric surface patch estimation.
• Support Vector Machine (SVM) classification for learning Perceptual Grouping.
• Graph-Cut optimization for object hypotheses generation.

Main Results:

• Successful segmentation of objects with arbitrary shapes in cluttered scenes.
• Effective handling of stacked, jumbled, and partially occluded objects.
• Demonstrated global optimality and prevention of incorrect groupings via Graph-Cut.

Conclusions:

• The proposed hierarchical framework significantly advances object segmentation in challenging scenarios.
• The integration of perceptual grouping and Graph-Cut offers a robust solution for complex scenes.
• The method shows competitive performance against state-of-the-art techniques on public datasets.