COCO Dataset


Learn more at
The COCO dataset has been developed for large-scale object detection, captioning, and segmentation. The 2017 version of the dataset consists of images, bounding boxes, and their labels
* Certain images from the train and val sets do not have annotations.
* Coco 2014 and 2017 datasets use the same image sets, but different train/val/test splits
* The test split does not have any annotations, only images.
* The dataset defines a total of 91 classes, but only uses 80.
* Panoptic annotations define 200 classes, but only uses 133.

COCO test-dev

Download Instructions

1) Go to:
2) Download the Train, Validation and Test Annotations and Images

COCO minival


Download Instructions

1) Go to:
2) Download the files available for Images and Annotations