The research focuses on pose estimation methods combining machine learning and optimization. The thesis introduces an efficient graph attention network model for learning structure-from-motion, which achieves accurate reconstructions. Techniques for enhancing the equivariance properties of convolutional neural network models are also presented, improving the accuracy of pose estimation. The thesis explores semidefinite relaxations of pose optimization problems, finding that absolute camera pose estimation may not always be solvable using these relaxations. Additionally, a rendering-based object pose refinement method and a long-term visual localization method are introduced. These advancements have the potential to enhance real-world applications in autonomous navigation, robotics, and augmented reality. Overall, this research contributes to the advancement of pose estimation techniques in computer vision.
Machine Learning Techniques Revolutionize Camera Pose Estimation
Date:
Updated: [falahcoin_post_modified_date]