COCO Dataset 2016

Description

• Microsoft COCO: Common Objects in Context
• Common Objects in Context

The challenge uses several metrics, but among typically reported are:

• Mean Average Precision
• Average Precision @ 0.5, 0.75 IoU for labels on bounding boxes having >= 0.5 IoU with ground truth — primary metric.
• AP-small — AP for instances with areas (total pixels) < 1024
• AP-medium — AP for instances with 1024 <= area < 9216
• AP-large — AP for instances with area >= 9216

Downloads

• Train images
• Validation images
• Train/val annotations
• Test images
• Test images info

Technical details

• Train set: ~83000 images, 604907 instances
• Validation set: ~41000 images
• 80 classes joined into 12 supercategories
• Number of pixels: 90% of images have between 167500 and 355840 pixels in train set.
• Height/width ratio: 90% of images have the ration between 0.6 and 1.5 in train set.
• Number of instances: 99% of images contain between 1 and 33 different instances, median = 4 (train set).

Format

Target labeling in contained in separate JSON files. Bounding boxes and corresponding labels are provided in instances_train2014.json and instances_val2014.json.

Instance annotations are stored in lists under the annotations key in corresponding files. Each object contains an annotation of one instance in one image. For the detection task the following fields are of particular interest:

• image_id: ID of the input image in the database
• bbox: 4-tuple containing the keypoint of a rectangular bounding box
• category_id: number of class associated with the instance

Information about categories, such as names and ids can be found under the key categories. Information about images (file name, resolution, id, area) can be found under the key images.

The maintainers of the dataset provide a multi-language API (listed below) for handling the index for the entire database.

Resources

Links to helper scripts for extraction/transformation/visualization.

Link	Description
COCO API	Lua, Matlab, and Python APIs for handling COCO dataset.

Demo