CORe50

CORe50，专为 (C) 连续 (O) 对象 (Re) 认知而设计，集合了 10 大类 50 种家居用品：插头适配器、手机、剪刀、灯泡、罐头、眼镜、球、记号笔、杯子、和遥控器。可以在对象级别（50 个类）或类别级别（10 个类）执行分类。第一个任务（默认）更具挑战性，因为在某些姿势中很难区分同一类的对象。

该数据集已在具有不同背景和照明的 11 个不同会话（8 个室内和 3 个室外）中收集。对于每个会话和每个对象，使用提供 300 个 RGB-D 帧的 Kinect 2.0 传感器录制 15 秒视频 (20 FPS)。

对象由操作员握住，摄像机视图是操作员眼睛的视点。操作员需要伸出手臂，平稳地移动/旋转摄像机前的物体。物体在抓取距离处的主观视角非常适合许多机器人应用。抓手（左或右）在整个会话期间发生变化，相关的对象遮挡通常由手本身生成。

CORe50, designed for (C)ontinuous (O)bject (Re)cognition, includes 50 household items categorized into 10 broad classes: plug adapters, mobile phones, scissors, light bulbs, cans, glasses, balls, markers, cups, and remote controls. Classification can be conducted at either the object level (50 classes) or the category level (10 classes). The first task (default setting) is more challenging, as distinguishing objects of the same class can be difficult under certain poses.

This dataset was collected in 11 distinct sessions (8 indoor and 3 outdoor) with varying backgrounds and lighting conditions. For each session and each object, 15-second videos (20 FPS) were recorded using a Kinect 2.0 sensor, which provides 300 RGB-D frames per video clip.

Objects are held by an operator, with the camera viewpoint matching the operator’s eye-level perspective. The operator is required to extend their arm and smoothly move or rotate the object in front of the camera. The subjective viewpoint of the object at grasping distance is highly suitable for numerous robotic applications. The grasping hand (left or right) changes throughout each session, and relevant object occlusions are often generated by the hand itself.

搜集汇总

背景与挑战

背景概述

CORe50是一个专为连续对象认知设计的图像数据集，包含10大类50种家居用品，可在对象或类别级别进行分类，其中对象级别分类更具挑战性。该数据集在11个不同会话中收集，使用Kinect 2.0传感器录制RGB-D视频，提供操作员主观视角，适用于机器人应用研究，于2017年由博洛尼亚大学发布。