Mgs-Stereo: Multi-Scale Geometric-Structure-Enhanced Stereo Matching for Complex Real-World Scenes.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Geometric Deep Learning

Background:

• Stereo matching is crucial for 3D reconstruction but struggles with complex real-world scenes.
• Challenges include non-Lambertian surfaces, textureless regions, and occlusions, hindering accurate pixel matching.
• Existing methods often fail to robustly handle geometric complexities in diverse environments.

Purpose of the Study:

• To develop a multi-scale stereo matching model that leverages scene geometry for improved accuracy.
• To enhance robustness in challenging regions like edges, occlusions, and reflective surfaces.
• To provide a more reliable stereo matching solution for complex, real-world imaging conditions.

Main Methods:

• A geometric structure perception module extracts scene geometric information.
• A geometric structure-adaptive embedding module fuses geometric and similarity features for disparity residual prediction.
• A geometric-based normalized disparity correction module refines matching in difficult areas.

Main Results:

• The proposed model achieves competitive performance on standard stereo matching benchmarks.
• Demonstrates robust and accurate disparity predictions in challenging scenarios.
• Outperforms leading approaches, particularly in regions with complex geometric and surface properties.

Conclusions:

• The multi-scale geometrically enhanced stereo matching model effectively addresses limitations of existing methods.
• Geometric information integration significantly improves robustness and accuracy in complex scenes.
• The model offers a promising advancement for real-world stereo vision applications.