CECMR: Calibrated Evidential Learning For Cross Modal Remote Sensing Image-Text Retrieval

Most existing cross modal remote sensing image-text retrieval (CMRSITR) methods overlook the importance of uncertainty estimation for test samples, which is critical for addressing the noisy and unpredictable nature of real-world environments. To tackle this issue, we propose a calibrated evidential learning based-method for CMRSITR (CECMR) that considers the reliability of retrieved samples. We use evidential learning (EDL) to model inter-modal correspondences and apply uncertainty consistency (UC) learning to ensure that the uncertainty serves as a meaningful indicator of reliability during inference. Additionally, we introduce mentor models with extensive modal-specific knowledge, helping our model refine and learn intra-modal relationships. Experimental results demonstrate that the proposed method outperforms existing state-of-the-art approaches and effectively evaluates test sample uncertainties.