VCR (Visual Commonsense Reasoning)

VCR的全称是视觉常识推理，是视觉常识推理的大规模数据集。这个数据集提出了关于图像的具有挑战性的问题，机器需要完成两个子任务: 正确回答问题和提供理由证明其答案。 VCR数据集包含大量问题，212K用于训练，26K用于验证，25K用于测试。答案和理由来自超过110K个独特的电影场景。VCR, short for Visual Commonsense Reasoning, is a large-scale dataset for visual commonsense reasoning. This dataset poses challenging questions related to images and requires machines to complete two subtasks: correctly answering the questions and providing justifications to support their answers. The VCR dataset contains a large quantity of questions, with 212K allocated for training, 26K for validation, and 25K for testing. Both the answers and justifications are sourced from over 110K unique movie scenes.

搜集汇总

背景与挑战

背景概述

VCR是一个大规模视觉常识推理数据集，旨在通过问题和理由证明两个子任务来评估机器理解能力。它包含约263K个样本，其中212K用于训练，数据来源于超过110K个电影场景。