Hierarchical Stereo Matching in Two-Scale Space for Cyber-Physical System

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Dense disparity map estimation from high-resolution stereo images presents significant challenges in matching accuracy and computational efficiency.
• Down-sampling high-resolution images risks losing crucial high-frequency components, leading to ambiguous correspondences and erroneous disparity estimates.
• Propagation of errors across scale space further exacerbates inaccuracies in disparity estimation.

Purpose of the Study:

• To develop an efficient hierarchical stereo matching method for high-resolution images.
• To enhance both matching accuracy and computational efficiency in dense disparity map estimation.
• To address the loss of high-frequency information during image down-sampling.

Main Methods:

• A two-scale space hierarchical stereo matching approach was implemented.
• Disparity estimation was performed on a reduced-resolution image and then up-sampled to the original resolution.
• High-frequency components (edges) extracted using Difference of Gaussians (DoG) or Canny operator were integrated, followed by edge-aware disparity propagation for refinement.

Main Results:

• The proposed algorithm demonstrated superior performance compared to existing methods.
• Accurate dense disparity maps were generated from high-resolution stereo images.
• Improved computational efficiency was achieved through the hierarchical processing.

Conclusions:

• The novel hierarchical stereo matching method effectively overcomes the limitations of traditional approaches for high-resolution images.
• The integration of high-frequency information and edge-aware refinement significantly enhances disparity map accuracy.
• This method offers a promising solution for accurate and efficient dense disparity estimation in computer vision applications.