COCO 2017

1. 简介该数据集是 COCO 系列的更新版本，在 COCO 2017 基础上调整了训练 / 验证集拆分，包含 118K 张训练图像（带检测、分割、关键点注释）、5K 张验证图像（带完整注释），以及 120K 张未标记图像（反映标记数据类分布）；训练集中 40K 图像和 5K 验证图像额外提供 “东西注释”，支持语义分割研究；测试集分为开发和挑战两个拆分，取代此前四类拆分，整体注释在检测、关键点维度与往年一致，新增 “东西注释” 维度This dataset is an updated iteration of the COCO series, with modifications to the training/validation set split based on COCO 2017. It includes 118K training images accompanied by detection, segmentation, and keypoint annotations, 5K validation images with complete annotations, and 120K unlabeled images that mirror the class distribution of the labeled data. Additionally, 40K images from the training set and the 5K validation images are provided with "stuff and thing annotations" to support semantic segmentation research. The test set is divided into two splits: development and challenge, replacing the previous four splits. The overall annotations are consistent with those of previous years for detection and keypoint tasks, with the new addition of the "stuff and thing annotations" dimension.

搜集汇总

背景与挑战

背景概述

COCO 2017是COCO数据集的更新版本，包含118K带注释的训练图像和5K验证图像，支持检测、分割和关键点任务，并新增“东西注释”以增强语义分割能力。它还包括120K未标记图像，适用于监督和半监督学习，广泛应用于计算机视觉多任务研究。