Fangming Zhong, Jing Wang, Tong Liu, Suhua Zhang, Zhikui Chen · ACM International Conference on Multimedia (ACM MM) 2026 · 2026
DOI: 10.1145/3767308.3835120
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Cross-modal retrieval with noisy labels has drawn significant research interest due to inevitable annotation noise during data collection. Several approaches have been proposed to tackle this challenge and have achieved promising success. However, they still suffer from the unreliability issue. On the one hand, noisy labels can mislead the model into overemphasizing non-discriminative features, resulting in significant errors in estimating noisy class posteriors. On the other hand, the label correction often relies on model predictions obtained from the softmax function, which can be overconfident and unreliable due to the translation invariance property of the conventional softmax. To tackle these challenges, we propose a framework entitled Dynamic Masking Regularization meets Evidential Calibration (DMREC) for cross-modal retrieval with noisy labels. Specifically, a Reliability-guided Dynamic feature Masking (RDM) scheme is designed to explicitly impose regularization, preventing the model from overfitting noisy samples. RDM adaptively regularizes the input data by masking the wrongly high-activated features and modulates the label simultaneously. In addition, an Evidential Prediction Calibration (EPC) module based on the Dirichlet distribution is introduced to break softmax's translation invariance. Extensive experiments demonstrate the significant superiority of our DMREC.
ACM International Conference on Multimedia (ACM MM) 2026 · 0 citations
No comments yet — start the discussion below.