MultiMNIST

MultiMNIST 数据集是从 MNIST 生成的。训练和测试是通过将一个数字叠加在来自同一组（训练或测试）但不同类别的另一个数字之上来生成的。每个数字在每个方向上最多移动 4 个像素，从而产生 36×36 的图像。考虑到 28×28 图像中的一个数字被限定在一个 20×20 的盒子中，两个数字边界框平均有 80% 的重叠。对于 MNIST 数据集中的每个数字，生成 1,000 个 MultiMNIST 示例，因此训练集大小为 60M，测试集大小为 10M。The MultiMNIST dataset is generated from the MNIST dataset. Training and test subsets are created by superimposing one digit onto another digit from the same subset (training or test) but belonging to a different class. Each digit is shifted by up to 4 pixels in each direction, resulting in 36×36 resolution images. Given that a single digit in the original 28×28 MNIST image is constrained within a 20×20 bounding box, the bounding boxes of the two digits have an average overlap of 80%. For each digit in the original MNIST dataset, 1,000 MultiMNIST examples are generated, leading to a training set size of 60M and a test set size of 10M.

搜集汇总

背景与挑战

背景概述

MultiMNIST是一个基于MNIST的大规模合成图像分类数据集，通过将两个不同类别的数字叠加并轻微位移生成36×36像素图像，训练集和测试集分别包含60M和10M样本。该数据集主要用于评估胶囊网络等模型的性能，特点是数字边界框平均有80%的重叠，增加了分类难度。