Multistage branch-and-bound merging for planar surface segmentation in disparity space

Area of Science:

• Computer Vision
• Image Processing
• Computational Geometry

Background:

• Stereo vision systems generate disparity images, which represent depth information.
• Accurate segmentation of planar surfaces in disparity images is crucial for 3D scene reconstruction and analysis.
• Existing methods may struggle with complex scenes or computational efficiency.

Purpose of the Study:

• To develop an iterative split-and-merge framework for segmenting planar surfaces in disparity space.
• To model scenes by approximating surfaces as planar and optimizing plane parameters.
• To efficiently estimate the number and labeling of planar surfaces in stereo images.

Main Methods:

• An iterative split-and-merge framework is proposed for disparity image segmentation.
• The framework models scene disparity by approximating surfaces as planar.
• A multistage branch-and-bound algorithm is employed for efficient optimization during merging.

Main Results:

• The framework iteratively refines planar surface segmentation by splitting and merging regions.
• The method minimizes residuals between actual and modeled disparity, optimizing plane parameters.
• Experimental results demonstrate the framework's effectiveness on various stereo image datasets.

Conclusions:

• The presented iterative split-and-merge framework provides an efficient method for planar surface segmentation in disparity images.
• The approach effectively models scenes with planar approximations and optimizes surface identification.
• The multistage branch-and-bound algorithm contributes to the computational efficiency of the segmentation process.