OOD-CV

OOD-CV数据集由约翰斯·霍普金斯大学创建，专注于评估视觉模型对自然图像中分布偏移的鲁棒性。该数据集包含10个对象类别的图像，涵盖姿态、形状、纹理、上下文和天气条件的变化。数据集通过互联网搜索和手动筛选收集，确保每张图像仅包含一个变量，同时其他变量与训练数据相似。OOD-CV数据集不仅用于图像分类，还支持对象检测和3D姿态估计任务，旨在推动对视觉模型鲁棒性的深入研究。The OOD-CV dataset, created by Johns Hopkins University, focuses on evaluating the robustness of visual models against distribution shifts in natural images. It contains images from 10 object categories, covering variations in pose, shape, texture, context, and weather conditions. Collected via internet searches and manual filtering, the dataset ensures that each image has only one shifted variable, while other variables are consistent with those in the training data. The OOD-CV dataset supports not only image classification tasks but also object detection and 3D pose estimation, aiming to promote in-depth research on the robustness of visual models.

搜集汇总

背景与挑战

背景概述

OOD-CV数据集由约翰斯·霍普金斯大学开发，旨在评估视觉模型对自然图像中分布偏移的鲁棒性，包含10个对象类别的图像，涵盖姿态、形状等多变量变化，并通过控制变量设计确保数据质量。该数据集支持图像分类、对象检测和3D姿态估计等任务，推动视觉模型鲁棒性的深入研究。