An application of stereo matching algorithm based on transfer learning on robots in multiple scenes.

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Binocular vision methods in robotics face challenges like high costs, complex algorithms, and unreliable disparity maps.
• Existing methods struggle with cross-domain generalization in diverse robotic scenes.

Purpose of the Study:

• To propose Ct-Net, a cross-domain stereo matching algorithm leveraging transfer learning for enhanced robot vision.
• To improve the reliability and efficiency of stereo matching in various robotic applications.

Main Methods:

• Ct-Net utilizes a General Feature Extractor and a feature adapter for domain adaptation.
• A Domain Adaptive Cost Optimization Module and disparity score prediction refine matching costs and search ranges.
• The framework employs a phased training strategy and ablation experiments for validation.

Main Results:

• Ct-Net significantly reduces 3PE-fg error on the KITTI 2015 benchmark (19.3% overall, 21.1% non-occluded).
• Achieved at least a 28.4% improvement in sample error rate for Staircase samples on the Middlebury dataset.
• Demonstrated superior cross-domain performance across Middlebury, Apollo, and real-world datasets.

Conclusions:

• Ct-Net effectively enhances cross-domain stereo matching performance for robot vision.
• The algorithm addresses limitations of current binocular vision methods, offering improved reliability and efficiency.
• Ct-Net shows practical applicability in diverse real-world robotic visual tasks.