Synthetic Dataset for Semantic Segmentation

该数据集由灰度图像组成，这些图像是通过将MNIST数据集中的数字放大并与前景和背景纹理结合生成的，旨在评估在语义分割任务中的上下文偏差。此外，数据集包含了有偏差和无偏差的版本，这允许我们比较不同显著性检测方法在探测上下文偏差方面的性能。所有图像的尺寸为64x64像素，该数据集的任务是进行带有上下文偏差检测的语义分割。

This dataset consists of grayscale images generated by upscaling handwritten digits from the MNIST dataset and combining them with foreground and background textures, designed to evaluate contextual bias in semantic segmentation tasks. Additionally, the dataset includes both biased and unbiased variants, enabling comparative assessment of the performance of different saliency detection methods in detecting contextual bias. All images have a resolution of 64×64 pixels, and the core task of this dataset is semantic segmentation with contextual bias detection.