Cost Volume Aggregation in Stereo Matching Revisited: A Disparity Classification Perspective

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Cost aggregation is crucial for stereo matching accuracy.
• Existing methods face challenges in efficient and informative cost aggregation.

Purpose of the Study:

• To propose a generic and efficient Disparity Context Aggregation (DCA) module.
• To improve Convolutional Neural Network (CNN)-based stereo matching performance.
• To leverage disparity classification for enhanced cost aggregation.

Main Methods:

• Classifying pixels into disparity classes to form homogeneous regions.
• Generating region representations to refine cost volumes.
• Integrating these representations into a shallow 3D CNN for cost aggregation.
• Developing a fully-differentiable DCA module compatible with various network architectures.

Main Results:

• DCA effectively suppresses irrelevant information and enhances matching.
• Homogeneous region representations enable efficient and informative cost aggregation.
• The DCA module improves performance with minimal overhead.
• A network incorporating DCA (DCANet) achieves state-of-the-art results on benchmarks.

Conclusions:

• Disparity class priors are beneficial for disparity regression in stereo matching.
• The proposed DCA module offers a significant advancement in cost aggregation techniques.
• DCANet demonstrates the practical effectiveness of the DCA module for high-accuracy stereo matching.