Project link: http://mscoco.org/

COCO is a dataset for image recognition, segmentation, and captioning sponsored by Microsoft. There are 80 object categories in this dataset with more than 300,000 images and two million instances. There are also challenges for detections, captions, and key-points every year.