DVD-Bench: A Benchmark for Dialogue-Centric Video Description

DVD-Bench is a benchmark for evaluating dialogue-centric video description, focusing on 'When, Who, and What is Said'. The dataset was created by tsinghua-ee and was last updated on 2026-06-24 10:56:52. It contains video files and corresponding annotation data in a parquet format for the English test split.

Use Cases

• Benchmarking video description models based on the 'When, Who, and What is Said' evaluation framework.
• Training models for temporal grounding of dialogue in videos based on the described annotation schema.
• Developing systems for speaker identification and attribution in video dialogues based on the benchmark's focus.

Strengths

• The dataset provides a specific benchmark for the focused task of dialogue-centric video description.
• The repository layout indicates a structured separation of video files and annotation data.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

tsinghua-ee

Freshness

Last updated 2026-06-24 10:56:52; freshness should be verified.