Explicit temporal inductive biases for video moment retrieval via multi-scale Gaussian priors

Guanlin Dai, Da Gao, Chao Zheng · 2026

Read Explicit temporal inductive biases for video moment retrieval via multi-scale Gaussian priors by Guanlin Dai, Da Gao, Chao Zheng (2026) 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/plos-pone-0354458.txt.

Description

Guanlin Dai, Da Gao, and Chao Zheng improve how AI finds a requested moment inside a video, a task called moment retrieval. Standard transformer models hunt the whole video at once because their search prompts carry no built-in sense of time span, and they falter on events of varied length. The team replaces those loose prompts with bell-shaped time priors that have a learnable center and width, seeded from local video features at several scales.

On a standard public benchmark, the design clearly helped one weak spot: finding long events, where scores rose by about two points over a leading rival with statistical support. Overall recall differences stayed inconclusive. Dai and colleagues present the prior as a cheap, focused part for studying scale-aware search in video models.

Catalog record

CREATORGuanlin Dai, Da Gao, Chao Zheng
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE59287
LICENSECC_BY
RIGHTSCC BY
SOURCEPLOS

Source: https://doi.org/10.1371/journal.pone.0354458

Browse the full library · About the Tower of Babel