Synthetic Visual Genome 2: 636,000 Panoptic Video Scene Graphs

Synthetic Visual Genome 2 (SVG2) provides spatio-temporal scene graphs for approximately 636,000 videos sourced from PVD and SA-V splits. Created by UWGZQ and released in 2026, the dataset integrates object labels, attributes, and instance-level segmentation masks for panoptic video analysis.

Use Cases

• Visual Question Answering using object labels and relationship predicates to reason about video content
• Video Scene Graph Generation by training on spatio-temporal relationship annotations
• Instance-level video segmentation utilizing the provided panoptic masks

Strengths

• Scale of ~636,000 videos across PVD and SA-V splits
• Multi-modal annotations including segmentation masks, attributes, and relationships
• Standardized Parquet format for efficient data processing

Limitations

• Synthetic data origin may result in a domain gap when applied to real-world video footage
• Temporal staleness may occur as the raw source data is not yet available for verification

Provenance

Source

UWGZQ, 'Synthetic Visual Genome 2: Extracting Large-scale Spatio-Temporal Scene Graphs from Videos'

Collection Method

Synthetic extraction and generation from video sources

Freshness

Last updated March 2026.