TDIUC 视觉问答数据集

TDIUC 全称 Task Directed Image Understanding Challenge，是一个视觉问答数据集。数据集包括来自 MS COCO 和视觉基因组数据集的 160 万个问题及 17 万张图像；image-question pair 被分为 12 个类别及 4 个额外的评价矩阵 (evaluation matrices)，用于评估模型稳健性及推理能力。

TDIUC, short for Task Directed Image Understanding Challenge, is a visual question answering (VQA) dataset. It comprises 1.6 million questions and 170,000 images derived from the MS COCO and Visual Genome datasets. The image-question pairs are categorized into 12 categories alongside 4 additional evaluation matrices, which are intended to assess the robustness and reasoning capabilities of models.

搜集汇总

背景与挑战

背景概述

TDIUC（Task Directed Image Understanding Challenge）是一个视觉问答数据集，包含来自MS COCO和视觉基因组的约167万张图像和165万个问题-答案对，问题被分为12个类别。该数据集还引入了4个新的评估指标和6个实验设置，旨在评估模型的稳健性和推理能力。