Cross-Modal Feature Matching

Cross-modal feature matching identifies which regions, points, or tokens in different modalities refer to the same scene element, enabling tasks such as image registration, 2D–3D alignment, retrieval, and sensor localization. Methods commonly learn shared or modality-adapted features using contrastive alignment, cross-attention, transformers, or pretrained vision models, then refine correspondences with geometric constraints, outlier rejection, and coarse-to-fine search. Evaluation spans settings such as RGB–infrared and optical–SAR registration, image-to-point-cloud matching on RGB-D Scenes and 7-Scenes, and camera–LiDAR localization on KITTI. A central challenge is balancing semantic similarity with geometrically valid matches while maintaining robustness across sensor shifts, domains, and deployment constraints.