During your thesis, you will research and develop advanced deep learning methods to improve keypoint matching accuracy for autonomous driving applications.
You will design a unified neural network architecture that jointly addresses correspondence refinement, outlier rejection, and uncertainty estimation, aiming to replace complex conventional post-processing chains.
Furthermore, you will investigate novel approaches to achieve sub-pixel precision and handle ambiguous image textures within the feature matching pipeline.
You will implement and train these models, focusing on efficiency and the potential for distilling the architecture into a fast, real-time capable solution.
Finally, you will benchmark your approach against current state-of-the-art matchers (e.g., in Visual Odometry scenarios) to demonstrate improvements in robustness and accuracy.