EgoExoBench is a benchmark designed to evaluate cross-perspective understanding in multimodal large models. It contains synchronized and asynchronous egocentric and exocentric video pairs with multiple-choice questions. The dataset was authored by Heleun and last updated on November 3, 2025.
Use Cases
- Benchmarking semantic alignment capabilities based on synchronized first- and third-person video pairs
- Evaluating viewpoint association reasoning based on cross-perspective video content
- Assessing temporal reasoning performance based on asynchronous video sequences
- Training models for cross-perspective video understanding based on the described question-answer format
Strengths
- Specifically designed for cross-perspective understanding, a distinct evaluation task
- Includes both synchronized and asynchronous video pairs, likely providing varied scenarios
- Last updated on 2025-11-03, suggesting recent maintenance
Limitations
- Column-level documentation is absent; field semantics must be inferred after download
- Row count is unknown, which may limit suitability assessment
- Description metadata is limited; actual data quality requires manual inspection after download
Provenance
- Source
- Heleun
- Collection Method
- Likely curated for benchmarking purposes, but specific collection method is not detailed.
- Time Range
- null
- Freshness
- Last updated 2025-11-03 13:53:29
- Geography
- null