Parallax attention stereo matching network based on the improved group-wise correlation stereo network.

Area of Science:

• Computer Vision
• Deep Learning
• Stereo Matching

Background:

• End-to-end deep stereo matching networks show promise in autonomous driving and depth sensing.
• Current state-of-the-art methods struggle with precise disparity prediction in near-range areas and at object edges.

Purpose of the Study:

• To enhance the precision of disparity prediction in stereo matching algorithms.
• To develop an end-to-end stereo matching method that accurately predicts both disparity and edge maps.

Main Methods:

• Proposed a parallax attention stereo matching algorithm integrated with an improved group-wise correlation stereo network.
• Introduced a novel parallax attention module operating at the three-dimensional (disparity, height, width) level.
• Developed a new edge detection branch and a multi-featured integration cost volume to leverage edge information.

Main Results:

• The parallax attention module improves feature expression, particularly in near-range regions, leading to high-precision disparity estimation.
• Integrating edge detection information enhances the accuracy of disparity estimation.
• The proposed method outperforms previous works on the Scene Flow and KITTI datasets.

Conclusions:

• The proposed parallax attention stereo matching algorithm significantly improves disparity prediction accuracy.
• The method's ability to learn from stereo correspondence and edge information makes it valuable for computer vision applications.
• This approach offers a robust solution for depth sensing and autonomous driving challenges.