AIC

首先，我们从dut中选择3402张图像来构建图像采集采样源，即ICSS，它涵盖72个类别，每个类别下至少有10张图像。随后，我们手动将72个类别分为11个主题 (表1)。ICSS中图像的纵横比范围从0.4625到1.9048。ICSS中的每个图像都标有类别，主题和显着性遮罩，因此从ICSS中采样的图像集合能够支持Mn和ms的计算。借助五重交叉验证的思想，我们将图像以4:1的比例划分为每个类别中的火车集和测试集，并且两个集合的分布几乎相同。

First, we selected 3,402 images from DUT to construct the Image Collection and Sampling Source, namely ICSS, which covers 72 categories with no fewer than 10 images per category. Subsequently, we manually divided these 72 categories into 11 themes (Table 1). The aspect ratios of images in ICSS range from 0.4625 to 1.9048. Each image in ICSS is annotated with its category, theme, and saliency mask, enabling the calculation of Mn and ms for image sets sampled from ICSS. Adopting the framework of 5-fold cross-validation, we split the images of each category into a training set and a test set at a ratio of 4:1, with the distributions of the two sets being nearly identical.

搜集汇总

背景与挑战

背景概述

AIC数据集基于DUT数据集构建，包含3402张图像，涵盖72个类别和11个主题，每张图像均标注了类别、主题和显著性遮罩。该数据集采用五重交叉验证方法划分为训练集和测试集，由清华大学与快手于2022年联合发布。