RKHS-BA: A Robust Correspondence-Free Multi-View Bundle Adjustment Framework for Semantic Point Clouds.

Area of Science:

• Robotics
• Computer Vision
• Geometric Deep Learning

Background:

• Bundle Adjustment (BA) is crucial for accurate pose estimation in 3D reconstruction.
• Traditional BA methods struggle with noisy data and diverse sensor inputs.
• Continuous representations offer potential for more robust and generalized landmark encoding.

Purpose of the Study:

• To develop a novel multi-frame Bundle Adjustment (BA) framework.
• To enable robust pose estimation using continuous landmark representations.
• To demonstrate generalized convergence beyond classical point-wise methods.

Main Methods:

• Introduced RKHS-BA, a framework utilizing Reproducing Kernel Hilbert Space (RKHS).
• Employed continuous landmark representations encoding RGB-D/LiDAR and semantic data.
• Utilized a correspondence-free pose graph formulation with a generalized loss function.

Main Results:

• Achieved highly robust pose estimations in extremely noisy scenes.
• Demonstrated strong generalization across various semantic inputs.
• Validated performance in multi-view point cloud registration, odometry, and LiDAR mapping.

Conclusions:

• RKHS-BA provides a generalized and robust approach to Bundle Adjustment.
• The framework shows significant improvements in challenging real-world scenarios.
• Continuous landmark representations in RKHS enhance pose estimation accuracy and reliability.