COCO

COCO (Common Objects in Context) is a widely used image dataset built to study object recognition in natural, cluttered scenes rather than isolated, iconic views. It contains over 330,000 images (more than 200,000 of them labeled) drawn largely from Flickr, featuring common objects in everyday contexts with multiple instances per image and substantial scale and occlusion variation.

What’s in it

COCO provides rich, layered annotations: about 1.5 million object instances with segmentation masks across 80 thing categories, additional stuff-segmentation labels for panoptic tasks, person keypoints for pose estimation, and five human-written captions per image. Standard train/val/test splits (notably the 2017 partition) and a public evaluation server make results directly comparable across methods and years. Derived tracks such as COCO Captions, COCO-Stuff, and the panoptic benchmark extended the original detection set into a broader vision-language and scene-parsing testbed.

Tasks, license & access

It is the standard benchmark for object detection, instance and panoptic segmentation, keypoint/pose estimation, and image captioning, with the widely cited mean Average Precision (mAP) protocol. Released by Microsoft, the annotations are distributed under a CC BY 4.0 license, while the underlying images remain subject to their original Flickr terms, so downloads are provided as image URLs plus annotation files.

• Modality: natural images with masks, boxes, keypoints, and captions
• Scale: 330K+ images, ~1.5M instances, 80 object categories
• Tasks: detection, segmentation, keypoints, captioning
• License: CC BY 4.0 annotations; images under Flickr terms
• Creator: Microsoft

COCO’s diverse, densely annotated scenes made it a default pretraining and evaluation set for vision models and a common bridge into multimodal vision-language systems.