EGO-VC – Evolutionary GPU-Optimization of Visual Correspondences for Image Alignment

Many computer vision applications, such as panoramic stitching, image morphing, and image registration depend on precise feature correspondences.While current machine-learning-based methods excel at detecting precise feature matches, they do not necessarily guarantee visual alignment.This work presents an evolutionary, meshbased optimization framework that enhances alignment quality and improves visual coherence as a postprocessing step for any feature-matching algorithm: The approach employs the well-known enhanced correlation coefficient (ECC) as a visual error metric, efficiently computed in parallel using the rasterization capabilities of modern graphics hardware.Given any state-of-the-art feature matcher, the proposed method constructs a Delaunay feature point mesh and evolutionarily refines the ECC alignment of each generated triangle.This results in a more precise registration beyond the accuracy of initial feature matches.Extensive evaluation across multiple standard image datasets confirms the proposed method's effectiveness, yielding ECC improvements up to 18.3% and ensuring high visual quality in downstream applications, particularly in challenging image areas with occlusions.For morphingbased applications, this leads to sharper, smoother transitions between image pairs, minimizing visual artifacts.