IconQA
收藏资源简介:
IconQA是一个大规模的数据集,包含107,439个问题,旨在评估抽象图标图像理解和视觉语言推理能力。该数据集由加州大学洛杉矶分校视觉、认知、学习和自主中心创建,包含三个子任务:多图像选择、多文本选择和填空。IconQA数据集灵感来源于现实世界的图表问题,强调了抽象图标理解的重要性,并要求模型不仅具备对象识别和文本理解等感知技能,还需要几何推理、常识推理和算术推理等多样化的认知推理技能。此外,为了帮助潜在的IconQA模型学习图标图像的语义表示,还发布了一个包含645,687个彩色图标的Icon645数据集,涵盖377个类别。
IconQA is a large-scale dataset consisting of 107,439 questions, designed to evaluate abstract icon image understanding and vision-language reasoning capabilities. This dataset was created by the Center for Vision, Cognition, Learning, and Autonomy at the University of California, Los Angeles (UCLA), and includes three subtasks: multiple-image selection, multiple-text selection, and fill-in-the-blank. Inspired by real-world chart-based questions, the IconQA dataset emphasizes the importance of abstract icon understanding, and requires models to possess not only perceptual skills such as object recognition and text understanding, but also diverse cognitive reasoning skills including geometric reasoning, commonsense reasoning, and arithmetic reasoning. Furthermore, to assist potential IconQA models in learning semantic representations of icon images, an additional dataset named Icon645 has been released, which contains 645,687 colored icons spanning 377 categories.

- 1IconQA: A New Benchmark for Abstract Diagram Understanding and Visual Language Reasoning加州大学洛杉矶分校视觉、认知、学习和自主中心 · 2022年



