Enhancing Video Moment Retrieval Using Training-Free IoU-Aware Temporal Clustering for Reranking

Recently, Video Moment Retrieval (VMR) has received significant increasing attention due to the growing demand for advanced video understanding. However, existing methods often rely on a single top-1 prediction, which frequently fails to capture semantically relevant moments, particularly in cases involving repetitive actions or temporally dispersed events. To address these limitations effectively, we propose a novel Training-free Clustering-based Reranking (TCR) method, which clusters top-ranked predictions based on temporal Intersection over Union (IoU) and selects representative predictions by jointly considering average IoU and confidence scores. Notably, our method does not require any additional training and achieves improved performance and efficiency on the Charades-STA dataset.