Accurate and Efficient Stereo Matching via Attention Concatenation Volume

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Stereo matching is crucial for 3D reconstruction and depth perception in vision and robotics.
• Efficient and accurate cost volume representation is key to advancing stereo matching algorithms.

Purpose of the Study:

• To introduce a novel cost volume construction method, Attention Concatenation Volume (ACV).
• To enhance stereo matching accuracy and efficiency by suppressing redundant information.
• To develop a real-time version (Fast-ACV) for practical applications.

Main Methods:

• Developed Attention Concatenation Volume (ACV) using attention weights from correlation clues.
• Integrated ACV into existing stereo matching networks for improved performance.
• Designed Fast-ACV using low-resolution correlation clues for real-time processing.
• Created ACVNet (high accuracy) and Fast-ACVNet (real-time) based on the proposed methods.

Main Results:

• ACV enables lighter aggregation networks while achieving higher accuracy in stereo matching.
• Fast-ACV significantly reduces computational and memory costs while maintaining satisfactory accuracy.
• ACVNet and Fast-ACVNet achieved state-of-the-art results on multiple benchmarks.

Conclusions:

• The proposed ACV method offers a significant advancement in cost volume representation for stereo matching.
• ACV and its fast variant provide versatile solutions for both high-accuracy and real-time stereo vision applications.