ADStereo: Efficient Stereo Matching With Adaptive Downsampling and Disparity Alignment.

Area of Science:

• Computer Vision
• Deep Learning
• Stereo Matching

Background:

• Balancing accuracy and computational efficiency is critical for deep learning stereo matching algorithms.
• Current methods often lose features during downsampling and suffer from spatial misalignment.
• Cost volume aggregation is a computationally intensive part of stereo matching.

Purpose of the Study:

• To develop novel sampling strategies for real-time stereo matching with high accuracy.
• To address feature loss and spatial misalignment issues in existing deep learning approaches.
• To introduce a new network, ADStereo, that prioritizes inference speed without compromising accuracy.

Main Methods:

• Introduced the Adaptive Downsampling Module (ADM) using local features and adaptive weights for effective downsampling.
• Developed the Disparity Alignment Module (DAM) with learnable interpolation to correct spatial misalignment.
• Integrated ADM and DAM into the ADStereo network architecture.

Main Results:

• ADStereo achieves real-time performance, outperforming state-of-the-art methods in speed.
• The network demonstrates comparable accuracy to existing leading stereo matching solutions on public benchmarks.
• ADStereo achieved 1.82% accuracy on KITTI 2015, faster than CREStereo.

Conclusions:

• The proposed ADM and DAM modules effectively enhance stereo matching accuracy and efficiency.
• ADStereo offers a superior balance between speed and accuracy for real-world stereo vision tasks.
• The developed methods provide a promising direction for efficient and accurate deep learning-based stereo matching.