Stereo matching based on multi-scale fusion and multi-type support regions.

Area of Science:

• Computer Vision
• Image Processing
• Stereo Vision

Background:

• Accurate disparity estimation is crucial for 3D reconstruction.
• Textureless regions pose significant challenges for traditional stereo matching algorithms.
• Existing methods often struggle to achieve high accuracy in uniform or repetitive areas.

Purpose of the Study:

• To develop a novel algorithm for robust disparity estimation in textureless and texture-free regions.
• To improve the accuracy and reliability of disparity maps generated from stereo images.
• To address the limitations of current state-of-the-art methods in challenging visual scenarios.

Main Methods:

• Fusion of color cost volume and gradient cost volume using a guided filter.
• Integration of multi-scale disparity maps guided by three types of image category information.
• Disparity refinement utilizing category-specific support regions and adaptive pixel weighting.

Main Results:

• The proposed algorithm demonstrates competitive performance compared to state-of-the-art methods.
• Effective handling of textureless regions, a known challenge in stereo vision.
• Accurate generation of primary disparity maps and refined disparity values.

Conclusions:

• The novel algorithm successfully addresses the challenge of disparity estimation in textureless regions.
• The integration of image category information significantly enhances disparity map accuracy.
• The method shows comparable or superior performance on benchmark datasets like Middlebury.