Visual grounding is the task of locating the image region that corresponds to a textual query. It typically requires large amounts of labeled data, which is expensive and difficult to curate at scale.

The paper, "Paint Outside the Box," investigates how to learn visual grounding when data is scarce. Rather than relying solely on manually collected datasets, the authors explore synthesizing training data and then selecting the most useful examples for learning.

The abstract is brief, so the specific synthesis and selection techniques are not detailed in the summary. Still, the work highlights a growing direction in multimodal learning: using synthetic data to ease annotation bottlenecks and improve model performance under limited supervision.