Sign in to view source links and access this dataset
Description
TimeLens2-93K is a large-scale dataset for temporal grounding in long videos, containing 23,793 videos and 93,232 text–temporal interval pairs. The dataset includes both single-span and multi-span evidence, with videos ranging from short clips to nearly 100 minutes in length. It was created by MCG-NJU and covers broad web domains such as entertainment, education, sports, news, science, gaming, travel, vehicles, music, and daily life.
Use Cases
Train models for temporal grounding based on text queries and video intervals.
Develop video retrieval systems based on the described broad web domains.
Benchmark multi-span evidence localization in long videos.
Analyze video content across diverse categories like entertainment, sports, and news.
Strengths
Large scale with 23,793 videos and 93,232 text-interval pairs.
Includes long-context videos up to nearly 100 minutes in duration.
Covers a broad range of 10+ web domains, suggesting diversity in content.
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
Row count is unknown, which may limit suitability assessment.
Data may reflect source bias inherent to web-scraped videos from huggingface.
Provenance
Source
MCG-NJU
Freshness
Last updated 2026-07-14 11:43:45; freshness should be verified.
License is unknown; terms of use must be verified before application.