OOD-CV-v2

OOD-CV-v2数据集是由爱丁堡大学等机构的研究人员创建，旨在评估视觉算法在真实世界场景中的鲁棒性。该数据集包含10个对象类别的分布外示例，涉及姿态、形状、纹理、上下文和天气条件的变化，支持图像分类、物体检测和3D姿态估计等多种视觉任务的模型基准测试。数据集通过互联网搜索和手动筛选收集，确保每个分布外数据样本仅在一个干扰因素上与训练数据不同，而其他因素保持相似。此外，数据集还增加了遮挡作为新的干扰因素，通过超现实遮挡模拟真实世界的遮挡情况。OOD-CV-v2数据集的构建和标注工作超过650小时，为研究分布外鲁棒性提供了丰富的测试平台，有望推动该领域的研究进展。

The OOD-CV-v2 dataset was developed by researchers from the University of Edinburgh and other institutions, aiming to evaluate the robustness of visual algorithms in real-world scenarios. This dataset includes out-of-distribution (OOD) examples across 10 object categories, covering variations in pose, shape, texture, context and weather conditions, and supports model benchmarking for multiple visual tasks such as image classification, object detection and 3D pose estimation. The dataset is collected through internet searches and manual screening, ensuring that each OOD sample differs from the training data only in one confounding factor, while other factors remain similar. Additionally, the dataset adds occlusion as a new confounding factor, simulating real-world occlusion scenarios via hyper-realistic occlusion methods. The construction and annotation work for the OOD-CV-v2 dataset totals over 650 hours, providing a rich testbed for research on out-of-distribution robustness and is expected to advance research progress in this field.

搜集汇总

背景与挑战

背景概述

OOD-CV-v2数据集由爱丁堡大学等机构开发，旨在评估视觉算法在真实世界场景中的分布外鲁棒性。它包含10个对象类别的分布外示例，涵盖姿态、形状、纹理、上下文、天气条件和遮挡等多种干扰因素，支持图像分类、物体检测和3D姿态估计等任务的基准测试，通过精心设计确保样本干扰因素单一化。