MultiVENT 2.0 is a benchmark for event-centric video retrieval containing over 218,000 news videos. It includes 108,500 videos for training and 109,800 for testing, with more than 3,900 queries targeting specific world events. The dataset is multilingual and was created by hltcoe.
Use Cases
- Train video retrieval models using the over 3,900 event-specific queries to match videos to world events.
- Benchmark multilingual event detection systems across the collection of 218,000+ news videos.
- Analyze the relationship between video content and textual event queries for cross-modal understanding.
Strengths
- Large scale with over 218,000 total videos.
- Structured for benchmarking with dedicated training (108,500 videos) and test (109,800 videos) splits.
- Multilingual focus supporting diverse event retrieval tasks.
Limitations
- Specific video metadata, features, or column structure are not described, limiting detailed analysis.
- The temporal and geographic coverage of the news events is unspecified, which may introduce bias.
- The dataset's size and format may require significant computational resources for processing.
Provenance
- Source
- hltcoe
- Collection Method
- null
- Time Range
- null
- Freshness
- null
- Geography
- null