Accuracy and efficiency stereo matching network with adaptive feature modulation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Feature enhancement is critical for robust matching tasks, especially in stereo vision where occlusion and reflection cause mismatches.
• Existing methods struggle with interference areas, leading to inaccurate predictions in challenging stereo matching scenarios.

Purpose of the Study:

• To propose an end-to-end dual-dimension feature modulation network (DFMNet) for enhanced feature representation in stereo matching.
• To develop a real-time variant, Fast-GFM, maintaining high accuracy and robustness.
• To introduce a cost filter volume (CFV) for adaptive filtering and improved matching accuracy.

Main Methods:

• DFMNet employs dual-dimension feature modulation (DFM) to capture spatial and channel information, adaptively combining local and contextual features.
• Cost Filter Volume (CFV) is introduced, using group-wise correlation weights to filter concatenated volumes and discard redundant information.
• Fast-GFM utilizes a global feature modulation (GFM) block for enhanced feature expression, enabling real-time stereo matching.

Main Results:

• DFMNet and Fast-GFM achieve state-of-the-art performance on benchmarks including Scene Flow, KITTI, ETH3D, and Middlebury.
• The proposed methods demonstrate significant improvements in matching accuracy and robustness across various stereo matching scenarios.
• Adaptive filtering via CFV effectively reduces redundant information, further enhancing stereo matching performance.

Conclusions:

• DFMNet and Fast-GFM offer effective solutions for improving feature representation and stereo matching accuracy.
• The dual-dimension feature modulation and cost filter volume approaches significantly enhance robustness in challenging conditions.
• The developed methods represent a substantial advancement in stereo matching technology, achieving top-tier results on multiple benchmarks.