openview2026/OpenView_data

OpenView数据集和基准是一个用于视觉问答（VQA）的数据集，专注于全景图像中的离视推理（OOV）。它包含超过158,000个高质量的多选OOV VQA问题，这些数据来自16,000个筛选的全景图像，覆盖11个主要场景类别。数据集采用多选VQA格式，每个问题包括问题类型、场景类别、室内/室外状态、一个透视投影视图的元数据、一个正确答案和四个干扰项，以及解释所有选项的理由。该数据集旨在为训练模型提供基于场景和上下文丰富的监督，以执行离视推理任务。数据集基于多个公开全景图像/视频数据集构建，提供OOV问答注释、源全景标识符和透视图参数，但原始全景图像或视频不重新分发。The OpenView Dataset and Benchmark is a dataset for visual question answering (VQA) focusing on out-of-view (OOV) reasoning in panoramic images. It contains over 158k high-quality multi-choice OOV VQAs derived from 16k filtered panoramic images, covering 11 major scene categories. The dataset uses a multi-choice VQA format, with each question including question type, scene category, outdoor/indoor status, metadata for one perspective-projected view, one correct answer and four distractors, and rationales explaining all options. It provides scene-grounded, context-rich supervision for training models to perform OOV reasoning. The dataset is built on public panoramic images/videos, offering OOV question-answer annotations, source panorama identifiers, and perspective-view parameters, but the original panorama images or videos are not redistributed.