DIS5K

为了构建高精度二分图像分割数据集 (DIS5K)，我们首先根据预先设计的关键字从Flickr1手动收集了12,000多个图像。然后，我们根据对象的结构复杂性从12,000图像中获得了22组225个类别的5,470个图像。然后使用GIMP以像素级精度手动标记每个图像。DIS5K中的标记目标主要集中在 “由预先设计的关键字 (类别) 定义的图像的对象” 上，而不考虑其特征，例如显着，常见，伪装，细致等。每张图像的平均标注时间为30分钟，有些图像的成本高达10小时。

To construct the high-precision binary image segmentation dataset (DIS5K), we first manually collected over 12,000 images from Flickr1 via pre-designed keywords. Then, we screened out 5,470 images belonging to 225 categories grouped into 22 sets from the initial 12,000 images based on the structural complexity of the target objects. Each image was subsequently manually annotated at pixel-level precision using GIMP. The annotation targets of DIS5K mainly focus on "the objects in images defined by pre-designed keywords (categories)", regardless of their inherent features such as saliency, commonness, camouflage, and meticulous details, etc. The average annotation time for each image is 30 minutes, while some images require up to 10 hours of annotation work.

搜集汇总

背景与挑战

背景概述

DIS5K是一个包含5,470个精细标注图像的高精度二分图像分割数据集，图像根据225个类别筛选并采用像素级手动标注，平均每张图像标注耗时30分钟。该数据集由苏黎世联邦理工学院等知名学术机构联合发布，旨在支持图像分割领域的研究。