Graph-based segmentation for RGB-D data using 3-D geometry enhanced superpixels.

Area of Science:

• Computer Vision
• 3-D Scene Understanding
• Image Segmentation

Background:

• RGB-D data is increasingly used for 3-D scene description due to advances in depth sensing.
• Accurate segmentation of RGB-D data is crucial for comprehensive scene analysis.

Purpose of the Study:

• To propose a novel two-stage segmentation method for RGB-D data.
• To improve 3-D scene segmentation by integrating color and depth information effectively.

Main Methods:

• Oversegmentation using 3-D geometry-enhanced superpixels with an 8-D distance metric.
• Graph-based merging of superpixels incorporating RGB-D proximity, texture, and boundary continuity.
• Label cost penalization for similar superpixels likely belonging to the same object.

Main Results:

• Qualitative and quantitative evaluations demonstrate the effectiveness of the proposed methods.
• The fusion of color and depth information leads to superior segmentation performance.
• The method outperforms several state-of-the-art RGB-D segmentation algorithms.

Conclusions:

• The proposed two-stage segmentation method effectively utilizes RGB-D data for accurate 3-D scene analysis.
• The integration of 3-D geometry in superpixel generation and graph-based merging significantly enhances segmentation quality.
• This approach offers a robust solution for semantic segmentation of complex 3-D environments.