PSG

PSG是一个经过精细注释的大规模数据集。数据集包含coco的48,749张图片，并基于coco已有的全景分割注解，标记了分割块之间的关系。数据集包含133个对象类别，其中80个是对象，53个是东西。
作者精细地定义了56种关系，包括位置关系 (上方，前方等)，常见的对象间关系 (悬挂等)，常见的生物动作 (行走，站立等)，人类行为 (烹饪等)，交通场景中的关系 (驾驶，骑马等)，运动场景中的关系 (踢腿等)，背景之间的关系 (包围等)。

PSG is a large-scale dataset with fine-grained annotations. It consists of 48,749 images sourced from the COCO dataset, and annotates the relationships between segmentation masks based on the existing panoptic segmentation annotations of COCO. The dataset covers 133 object categories, with 80 being thing categories and 53 being stuff categories.
The authors meticulously defined 56 types of relationships, including positional relationships (e.g., above, in front of), common inter-object relationships (e.g., hanging), common biological actions (e.g., walking, standing), human behaviors (e.g., cooking), relationships in traffic scenarios (e.g., driving, riding a horse), relationships in motion scenarios (e.g., kicking), and relationships between background regions (e.g., surrounding).

搜集汇总

背景与挑战

背景概述

PSG是一个基于COCO的大规模精细注释数据集，包含48,749张图片，133个对象类别和56种关系类型，由商汤科技和南洋理工大学实验室于2022年发布。