CLEVR-Hans

CLEVR-Hans 数据集是一种新颖的混杂视觉场景数据集，它捕获了不同对象的复杂组合。该数据集由分为几类的 CLEVR 图像组成。
类的成员资格基于对象的属性和关系的组合。此外，数据集中的某些类是混淆的。因此，在由训练、验证和测试分割组成的数据集中，所有混淆类的训练和验证图像都将与特定属性或属性组合混淆。
每个类由 3000 个训练图像、750 个验证图像和 750 个测试图像表示。对于 CLEVR-Hans3，训练、验证和测试集拆分分别包含 9000、2250 和 2250 个样本，对于 CLEVR-Hans7，分别包含 21000、5250 和 5250 个样本。所有数据拆分的类分布都是平衡的。
对于类规则包含三个以上对象的 CLEVR-Hans 类，每个场景要放置的对象数量是在该类所需的最小对象数量和十个之间随机选择的，而不是在三到十之间，如在原始 CLEVR 数据集。
最后，创建的图像使得类规则的精确组合不会出现在其他类的图像中。一个类规则的对象子集可能出现在另一类的图像中。但是，图像中不可能包含多个完整的类规则。

CLEVR-Hans is a novel confounding visual scene dataset that captures complex combinations of distinct objects. This dataset consists of CLEVR images grouped into several categories. Class membership is based on a combination of object attributes and relationships. Furthermore, certain classes within the dataset are confounding. Therefore, in the dataset composed of training, validation, and test splits, all training and validation images of confounding classes will be confounded with specific attributes or attribute combinations.

Each class is represented by 3000 training images, 750 validation images, and 750 test images. For CLEVR-Hans3, the training, validation, and test splits contain 9000, 2250, and 2250 samples respectively; for CLEVR-Hans7, the splits contain 21000, 5250, and 5250 samples respectively. The class distribution is balanced across all data splits.

For CLEVR-Hans classes whose class rules require more than three objects, the number of objects to place in each scene is randomly selected between the minimum number of objects required for that class and ten, rather than between three and ten as in the original CLEVR dataset.

Finally, the generated images ensure that the exact combination of a class rule does not appear in images of other classes. A subset of objects from one class rule may appear in images of another class, but it is impossible for an image to contain multiple complete class rules.

搜集汇总

背景与挑战

背景概述

CLEVR-Hans是一个混杂视觉场景数据集，包含不同对象的复杂组合，分为训练、验证和测试集，样本数量在不同版本中有所不同，且类分布平衡。数据集中的某些类是混淆的，训练和验证图像将与特定属性或属性组合混淆。