PSG: Panoptic Scene Graph Benchmark in COCO Format with Bounding Boxes

Panoptic Scene Graph (PSG) benchmark reformatted into COCO-JSON by maelic for the SGG-Benchmark framework. It contains between 10,000 and 100,000 records derived from the 2022 NeurIPS publication by Yang et al. and was utilized for the 2025 REACT paper. The data is optimized for visual relationship detection and object detection pipelines.

Use Cases

• Visual relationship detection using scene graph labels to map interactions between objects
• Object detection training using the bounding box coordinates provided in the COCO-JSON structure
• Scene graph generation benchmarking against the REACT paper (BMVC 2025) results

Strengths

• Standardized COCO-JSON format compatible with common computer vision frameworks
• Scale of 10,000 to 100,000 annotated records
• MIT licensed for open research and commercial use

Limitations

• Lacks original panoptic segmentation masks present in the source PSG dataset
• Restricted to bounding box geometry for object localization

Provenance

Source

Yang et al., NeurIPS 2022 (Panoptic Scene Graph)

Collection Method

Reformatting of existing benchmark data into COCO-JSON standard

Time Range

2022-2025

Freshness

Last updated March 2026.