Dense Correspondence

Dense correspondence, the task of establishing pixel-level or point-level matches between images or between images and 3D point clouds, is crucial for numerous computer vision applications. Current research focuses on improving the accuracy and robustness of correspondence estimation, particularly in challenging scenarios like large-scale scenes, significant viewpoint changes, and non-rigid deformations, employing techniques such as transformer networks, diffusion models, and optimal transport. These advancements are driving progress in diverse fields, including Structure from Motion, multi-camera object tracking, 3D reconstruction, and robotic perception, enabling more accurate and reliable solutions for tasks such as object pose estimation, scene understanding, and autonomous navigation.