DBCON: Dual Bias Control in Zero-Shot Video Moment Retrieval

Zero-shot Video Moment Retrieval (ZVMR) is a task that localizes temporal moments within a video using only natural language queries, without requiring large-scale datasets or extensive training resources. While previous ZVMR studies have focused on aligning high-quality video and query representations, they have faced performance limitations by overlooking the ‘Dual Bias’ problem, which arises from the excessive specificity of queries and the generality of video captions. To address this challenge, we propose the DBCON (Dual Bias CONtrol) framework, which simultaneously controls caption specificity and query generalization. DBCON employs ‘Query-Aware Caption Refinement’ to induce captions that explicitly reflect the core information of the query, and ‘Query Debiasing’ to rephrase the user query into a more universal expression via an LLM. Our experiments show that DBCON effectively mitigates this dual bias, achieving up to 1.73% higher retrieval accuracy on mAP@avg compared to the state-of-the-art model on the QVHighlights test benchmark. This study is significant as it presents a new direction for developing scalable and efficient video-language retrieval systems.