OMTG Benchmark Exposes MLLMs' Failure to Localize Multiple Video Segments per Query

One-to-Many Temporal Grounding (OMTG) formalizes the task of localizing multiple disjoint video segments for a single query, introducing Count Accuracy and Effective Temporal F1 metrics alongside a 56K-sample dataset. State-of-the-art MLLMs tuned for single-segment retrieval score near zero on this benchmark due to missing event cardinality perception.