COCO (Common Objects in Context)
COCO Consortium (Microsoft and academic partners)
330,000 everyday photos with every person, dog, and fork outlined by hand.
Images of everyday scenes with pixel-level outlines of objects in 80 categories, five human-written captions per image, and body keypoints for people. It is the standard benchmark for object detection, segmentation, and image captioning, and most detection models report COCO scores.
Access
- Cost
- FreeFree download.
- License
- Open licenseAnnotations under CC BY 4.0; images remain under their Flickr licenses
- Sign-up
- No sign-up
- How you get it
- API and Bulk downloadJPEG, JSON
What's in it
- One row is
- One image with its object instances, captions, and person keypoints
- Covers
- Flickr photos, worldwide
- Years
- Released 2014; 2017 split is the current standard
- Updated
- One-time release
- Size
- About 330,000 images, 1.5 million object instances, 80 object categories
Quality
Carefully annotated, versioned, and supported by an evaluation API; the default benchmark for detection.
- Only 80 object categories
- Images carry a mix of Flickr licenses, some non-commercial
- Captions and people skew toward Western contexts
Connects to
- ImageNet: Classification counterpart; models are often pretrained on ImageNet and fine-tuned on COCO
- Open Images V7: A larger detection set with 600 box classes
Questions it could answer
- Which object categories co-occur most often in everyday scenes?
- How does caption wording differ for images with and without people?