FunnyBirds
收藏资源简介:
FunnyBirds是一个用于可解释AI方法部分分析的合成视觉数据集,由达姆施塔特工业大学计算机科学系创建。该数据集包含50,500张图像,分为50,000张训练图像和500张测试图像,涵盖50种合成鸟类物种。数据集设计强调概念,即人类可理解的心理实体,如鸟的喙、翅膀、脚、眼睛和尾巴。创建过程中,通过添加类别特定的部分到中性身体模型来生成每个FunnyBird实例,并随机添加背景对象和改变光照及视角以反映现实世界的挑战。FunnyBirds数据集旨在通过允许进行语义上有意义的图像干预,如移除单个对象部分,来解决可解释AI领域中自动评估未解决问题。该数据集的应用领域包括分析和评估不同解释类型的可解释AI方法,以及揭示现有方法的强弱点,从而推动可解释AI的发展。
FunnyBirds is a synthetic visual dataset for partial analysis of explainable AI (XAI) methods, created by the Department of Computer Science at Technische Universität Darmstadt. It contains 50,500 images, split into 50,000 training images and 500 test images, covering 50 synthetic bird species. The dataset is designed to emphasize concepts, i.e., human-understandable mental entities such as a bird’s beak, wings, feet, eyes, and tail. During its creation, each FunnyBird instance is generated by adding category-specific parts to a neutral body model, with background objects randomly added and lighting and viewpoint varied to reflect real-world challenges. The FunnyBirds dataset aims to address the unsolved problem of automatic evaluation in the field of explainable AI by enabling semantically meaningful image interventions such as removing individual object parts. Its application scenarios include analyzing and evaluating explainable AI methods across different explanation types, as well as revealing the strengths and weaknesses of existing methods, thereby advancing the development of explainable AI.

- 1FunnyBirds: A Synthetic Vision Dataset for a Part-Based Analysis of Explainable AI Methods达姆施塔特工业大学计算机科学系 · 2023年



