Open Images V7
About 9 million Creative Commons photos with 16 million boxes and 61 million labels.
Flickr images released under Creative Commons, annotated with image-level labels across 20,000+ classes, bounding boxes for 600 object classes, segmentation masks, visual relationships, point labels, and localized narratives in which annotators spoke while tracing objects. It is the largest openly licensed detection set from a single provider.
Access
- Cost
- FreeFree download.
- License
- Open licenseAnnotations under CC BY 4.0; images listed as CC BY 2.0
- Sign-up
- No sign-up
- How you get it
- Bulk downloadCSV, JPEG, PNG masks
What's in it
- One row is
- One image with its labels, boxes, masks, relationships, and narratives
- Covers
- Flickr photos worldwide
- Years
- First released 2016; V7 in 2022
- Updated
- Irregular
- Size
- About 9 million images; 15.85 million boxes; 2.79 million masks; 61.4 million image-level labels
Quality
Large and permissively licensed with human-verified labels, though many image-level labels are machine-generated and only partly verified.
- Not every object in an image is boxed
- Image-level labels mix human-verified and machine-generated
- Photo-sharing bias toward certain regions and subjects
Connects to
- COCO (Common Objects in Context): Smaller, more exhaustively annotated benchmark for the same tasks
- ImageNet: Classification set with a different label hierarchy
Questions it could answer
- Which of the 600 box classes are rarest, and how does that affect detection accuracy?
- What do people mention first when narrating a scene?