OOD-CV

我们的ood-cv基准测试的目标是诊断计算机视觉模型的鲁棒性，以解决数据中的分布外变化。为了实现这一目标，基准测试由PASCAL VOC 2012和ImageNet数据集中的10个对象类别 (飞机，公共汽车，汽车，火车，船，自行车，飞机，椅子，餐桌，沙发) 组成。请注意，只有我们提供的训练数据才能用于训练模型，并且不允许使用外部训练数据。此限制使我们能够设计测试集，以使每个测试示例在一个特定的麻烦w.r.t中受到分布外的转移。训练数据，例如对象的3D姿势，形状，纹理，上下文，天气和遮挡。

Our OOD-CV benchmark aims to diagnose the robustness of computer vision models against out-of-distribution (OOD) shifts in data. To achieve this goal, the benchmark consists of 10 object categories selected from the PASCAL VOC 2012 and ImageNet datasets: airplane, bus, car, train, ship, bicycle, airplane, chair, dining table, sofa. Please note that only the training data provided by us can be used for model training, and the use of external training data is prohibited. This constraint allows us to design the test set such that each test example exhibits out-of-distribution shifts with respect to a specific aspect of the training data, such as the 3D pose, shape, texture, context, weather conditions, and occlusion of the objects.

搜集汇总

背景与挑战

背景概述

OOD-CV是一个用于评估计算机视觉模型对分布外变化鲁棒性的基准测试数据集，基于PASCAL VOC 2012和ImageNet的10个对象类别构建。该数据集由多所知名大学联合发布，旨在通过限制训练数据使用来专门测试模型在不同分布外条件下的表现。