BSEF-Stereo: A Stereo Matching Model Based on Branching Strategy and Error Feedback.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Stereo matching is difficult in areas with few textures, repetitive patterns, occlusions, and object edges.
• Current methods struggle to balance local detail preservation with broad contextual reasoning.

Purpose of the Study:

• To introduce BSEF-Stereo, an iterative framework for enhanced stereo matching.
• To address limitations of existing recurrent updaters in handling complex scene elements.

Main Methods:

• Utilizes a channel-position attention module to enhance spatial and channel cues.
• Employs a gated recurrent unit with complementary branches for detail preservation and context aggregation.
• Incorporates an error-aware refinement module using reprojection error and image guidance.

Main Results:

• Achieved competitive performance on Scene Flow, KITTI 2012, KITTI 2015, and Middlebury datasets.
• BSEF-Stereo obtained 0.41 px EPE and 2.27% D1 on Scene Flow.
• Reached a D1-all error of 1.48% on KITTI 2015.

Conclusions:

• BSEF-Stereo effectively combines adaptive updating and error feedback for robust stereo matching.
• The proposed modules contribute significantly to performance across diverse scenes.
• The framework demonstrates state-of-the-art results in challenging stereo vision tasks.