Focus : Efficient Keyframe Selection for Long Video Understanding

Zirui Zhu, Hailun Xu, Yang Luo, Yong Liu, Kanchan Sarkar Zhenheng Yang Yang You · 2025

Read Focus : Efficient Keyframe Selection for Long Video Understanding by Zirui Zhu, Hailun Xu, Yang Luo, Yong Liu, Kanchan Sarkar Zhenheng Yang Yang You (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-27280.txt.

Description

Zirui Zhu leads six co-authors in FOCUS, a keyframe picker that lets video models answer questions about hour-long footage within a tight token budget. Multimodal models turn frames into tokens, so long videos explode the budget. Pipelines either subsample uniformly or pre-filter with retrieval scoring, which can miss the most informative moments before selection even starts.

FOCUS treats short clips as bandit arms in a pure-exploration problem, using empirical means and a Bernstein confidence radius to find valuable regions while still exploring uncertain ones. A two-stage procedure first locates high-value regions, then picks top frames inside each. On two long-video benchmarks it improves accuracy substantially while processing under two percent of frames, with especially strong gains on videos over twenty minutes.

Catalog record

CREATORZirui Zhu, Hailun Xu, Yang Luo, Yong Liu, Kanchan Sarkar Zhenheng Yang Yang You
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE44 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-27280

Browse the full library · About the Tower of Babel