A Straightforward Framework for Video Retrieval Using CLIP

Video Retrieval is a challenging task where the aims at matching text query to video or vice versa. Most of existing approaches for addressing such problem rely on annotations made by users. Although simple, this approach not always feasible in practice. In work, we explore application language-image model, CLIP, obtain representations without need said annotations. This model was explicitly trained learn common space images and can be compared. Using various techniques described document, extended its videos, obtaining state-of-the-art results MSR-VTT MSVD benchmarks.