Video Question Grounding (VideoQG) is a task that requires both answering a question about a video and identifying the exact temporal segment that supports the answer. According to the abstract, this dual requirement makes the task particularly challenging.
Existing supervised methods for VideoQG rely on expensive temporal annotations, which limits their practicality. The abstract also mentions existing zero-shot approaches, though the excerpt provided cuts off before detailing their specific shortcomings.
The proposed method, DiffGround, is described as a semantic difference-guided zero-shot approach. The title suggests that it uses differences in semantic content to guide grounding, potentially avoiding the need for costly annotations.