SnAG: Scalable and Accurate Video Grounding

Temporal grounding of text descriptions in videos is a central problem invision-language learning and video understanding. Existing methods oftenprioritize accuracy over scalability – they have been optimized for groundingonly a few text queries within short videos, and fail to scale up to longvideos with hundreds of queries. In this paper, we study the effect ofcross-modal fusion on the scalability of video grounding models. Our analysisestablishes late fusion as a more cost-effective fusion scheme for long-formvideos with many text queries. Moreover, it leads us to a novel, video-centricsampling scheme for efficient training. Based on these findings, we presentSnAG, a simple baseline for scalable and accurate video grounding. Withoutbells and whistles, SnAG is 43state of the art for long-form video grounding on the challenging MAD dataset,while achieving highly competitive results on short videos.