Generating LLM-written start/end descriptions of actions and converting rigid boundary labels into similarity-weighted probability targets yields consistent but small gains in language-driven action localization across three datasets.
ArXiv preprint
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
LLM-powered Query Expansion for Enhancing Boundary Prediction in Language-driven Action Localization
Generating LLM-written start/end descriptions of actions and converting rigid boundary labels into similarity-weighted probability targets yields consistent but small gains in language-driven action localization across three datasets.