Programs

Correspondence Problem

This program focuses on the fundamental task of finding correspondences between parts of a scene in different images, a core problem in computer vision. It covers techniques for detecting distinctive local features in images that are invariant to geometric and photometric transformations. The program then explores how to establish tentative matches between these features and how to use robust algorithms like RANSAC to identify and verify accurate correspondences, filtering out incorrect ones based on a geometric model of the scene. The curriculum involves implementing these components from scratch using libraries like PyTorch and NumPy. Key areas include feature detection methods such as Harris detectors, automatic scale estimation using scale-space representations, computing local invariant descriptions, and establishing tentative correspondences followed by RANSAC for verification. The goal is to build a solid understanding of the underlying principles and practical implementation of correspondence finding in computer vision.