Openimages Narratives V2: 1.9 Million Images with Bounding Boxes and Captions

1.9 million images from the Open Images V7 project annotated with bounding boxes and captions. The annotations were generated automatically using image-text-to-text models. The dataset was uploaded by Fhrozen to Hugging Face and last updated on November 17, 2025.

Use Cases

• Train object detection models based on the bounding box annotations.
• Develop or fine-tune image captioning models based on the provided captions.
• Benchmark multimodal AI models using the paired image-text data.
• Perform data analysis on large-scale image annotation patterns.

Strengths

• 1.9 million images provide a substantial scale for model training.
• Annotations include both bounding boxes and captions, offering multiple supervision signals.

Limitations

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

Provenance

Source

Google Localized Narrative (original source for Open Images Dataset V7).

Collection Method

A subset of the original 9M images was annotated with automatic methods (image-text-to-text models).

Freshness

Last updated 2025-11-17 23:19:20; freshness should be verified.