LSE-CVCNet: A Generalized Stereoscopic Matching Network Based on Local Structural Entropy and Multi-Scale Fusion

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Dynamic scenes present challenges for stereo matching due to feature misalignment and occlusions.
• Existing methods struggle with texture variability and contextual ambiguity in complex environments.

Purpose of the Study:

• To develop a novel stereo matching network, LSE-CVCNet, for enhanced performance in dynamic scenes.
• To address challenges of feature misalignment and contextual ambiguity in disparity estimation.

Main Methods:

• Introduced Local Structural Entropy (LSE) for adaptive attention guided by disparity map uncertainty.
• Developed a Cross-Image Attention Mechanism (CIAM-T) for asymmetric feature extraction and improved alignment.
• Implemented Multi-Resolution Cost Volume Fusion (MRCV-F) for preserving fine-grained details across scales.

Main Results:

• LSE-CVCNet demonstrated superior disparity estimation accuracy and cross-domain generalization.
• The network showed robustness under varying lighting, occlusions, and complex geometries.
• Outperformed state-of-the-art methods on multiple datasets, validated by ablation and cross-domain tests.

Conclusions:

• LSE-CVCNet establishes a new paradigm for adaptive stereo matching in dynamic environments.
• The integrated modules effectively enhance feature alignment and detail preservation.
• The proposed method shows significant potential for real-world applications requiring robust stereo vision.