TartanMatch: Towards Universal Dense Correspondence Across Modalities

TartanMatch is a unified model for dense correspondence across all 25 ordered pairs of RGB, event, thermal, depth, and projected LiDAR modalities. It adapts an RGB-pretrained model using lightweight adaptation and real/synthetic data, achieving 61.9% and 49.6% reductions in Endpoint Error over baselines for cross-modal and same-modal matching, respectively. Joint training across modalities improves accuracy, especially for challenging cross-modal pairs.