3D Geometry-Aware Efficient Feature Matching for Weakly Textured Scenes

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Local feature matching is crucial for robotic SLAM and visual localization.
• Weakly textured industrial environments pose challenges for traditional appearance-based methods.
• Existing lightweight methods struggle with discriminative and stable feature learning in these settings.

Purpose of the Study:

• To propose GAEFeat, a novel lightweight vision-geometric feature learning network.
• To address the lack of specialized training data for feature learning in industrial environments.
• To enhance the robustness and efficiency of local feature matching in challenging scenes.

Main Methods:

• Developed a lightweight vision-geometric feature learning network named GAEFeat.
• Integrated robotic arm pose priors and depth information for automated cross-view supervision and surface-normal labels.
• Designed a dual enhancement mechanism with a geometric auxiliary branch and a geometry-aware enhancement (GAE) module.
• Constructed simulated and real-world datasets for training and evaluation.

Main Results:

• GAEFeat demonstrates strong robustness and high inference efficiency in relative pose estimation, homography estimation, and visual localization.
• Achieved notable advantages in near-field, weakly textured industrial scenes.
• Inference latency of 3.9 ms on NVIDIA Jetson AGX Orin, indicating real-time capability.

Conclusions:

• GAEFeat effectively learns discriminative and stable local features for weakly textured industrial environments.
• The proposed method offers practical potential for deployment in edge computing environments.
• The integration of geometric priors significantly improves feature matching performance and efficiency.