A CLIP-Hitchhiker's Guide to Long Video Retrieval

TL;DR

This paper adapts CLIP for long video retrieval by introducing a weighted-mean temporal aggregation method based on query-scoring, achieving state-of-the-art results with a simple yet effective approach.

Contribution

It proposes a novel weighted-mean temporal aggregation method for CLIP-based video retrieval, outperforming previous temporal modeling techniques.

Findings

Weighted-mean aggregation significantly improves retrieval performance.

The simple baseline outperforms complex temporal modeling methods.

Achieves state-of-the-art results on multiple long video retrieval benchmarks.

Abstract

Our goal in this paper is the adaptation of image-text models for long video retrieval. Recent works have demonstrated state-of-the-art performance in video retrieval by adopting CLIP, effectively hitchhiking on the image-text representation for video tasks. However, there has been limited success in learning temporal aggregation that outperform mean-pooling the image-level representations extracted per frame by CLIP. We find that the simple yet effective baseline of weighted-mean of frame embeddings via query-scoring is a significant improvement above all prior temporal modelling attempts and mean-pooling. In doing so, we provide an improved baseline for others to compare to and demonstrate state-of-the-art performance of this simple baseline on a suite of long video retrieval benchmarks.