COCO (Common Objects in Context) is one of the most classic benchmark datasets in the field of computer vision, containing 330,000 images, 1.5 million object instances, 80 object categories, and supporting tasks such as object detection, instance segmentation, keypoint detection, and image captioning.
The most influential benchmark dataset in the field of computer vision, covering various core visual tasks
Provides bounding box annotations for 80 object categories, covering common objects in daily life, and serves as the standard training and evaluation dataset for detection models such as YOLO and Faster R-CNN.
Each object instance comes with pixel-level segmentation masks, supporting fine-grained instance segmentation tasks, widely used for training and evaluation of models like Mask R-CNN.
Includes over 250,000 annotated human instances with 17 keypoints, serving as a core data source for human pose estimation and action recognition research.
Each image is accompanied by 5 human-written natural language descriptions, supporting tasks such as image captioning and visual question answering (VQA).
In addition to countable "thing" categories, 91 "stuff" categories (such as sky, grass, road) are also annotated, supporting panoptic segmentation.
Created under the leadership of Microsoft Research, it is the most cited visual dataset in top conference papers such as CVPR, ECCV, and ICCV, recognized as a benchmark in both academia and industry.
From academic research to industrial deployment, covering core tasks in computer vision
Train and evaluate object detection models, such as classic and cutting-edge architectures like YOLO, SSD, Faster R-CNN, DETR, etc.
A standard evaluation platform for instance segmentation, semantic segmentation, and panoptic segmentation tasks, supporting models like Mask R-CNN, SAM, etc.
Pose estimation research based on 17 human keypoint annotations, suitable for training models like OpenPose, HRNet, etc.
Train image description generation models using image-text paired data, supporting multimodal understanding and visual language pre-training.
The following is a JSON example of a single object instance from the COCO annotation file
{
"info": {
"description": "COCO 2017 Dataset",
"version": "1.0",
"year": 2017,
"contributor": "COCO Consortium",
"url": "http://cocodataset.org"
},
"images": [
{
"id": 397133,
"file_name": "000000397133.jpg",
"width": 640,
"height": 480
}
],
"annotations": [
{
"id": 1768,
"image_id": 397133,
"category_id": 18,
"bbox": [217.62, 240.54, 38.99, 57.75],
"area": 2254.6,
"segmentation": [[...]],
"iscrowd": 0
}
],
"categories": [
{"id": 1, "name": "person", "supercategory": "person"},
{"id": 18, "name": "dog", "supercategory": "animal"}
]
}
From browsing to loading, you can start your computer vision project in just a few minutes
View dataset details on the Ace Data Cloud platform, including category distribution, annotation format, and licensing agreements.
Download the training set (118K images, 18 GB), validation set (5K images, 1 GB), and annotation files, selecting the required subsets as needed.
Use pycocotools to load annotation data, quickly completing data visualization, model training, and evaluation processes.
The most classic benchmark dataset in the field of computer vision, open license, available for immediate download. Whether you are a deep learning researcher or an industrial application developer, COCO is an indispensable data resource.