RAVEN
收藏资源简介:
RAVEN数据集是由加州大学洛杉矶分校视觉、认知、学习和自主中心创建,旨在通过视觉与结构、关系和类比推理的关联,提升机器智能。该数据集包含70,000个Raven's Progressive Matrices问题,均分布在7种不同的图形配置中。每个问题有16个树结构注释,总计1,120,000个结构标签。数据集设计注重推理而非视觉识别,每张图像仅包含简单的灰度对象,规则按行应用,挑战机器在短期记忆和组合推理方面的弱点。RAVEN数据集通过引入结构表示,为机器提供了一种新的抽象推理方式,旨在推动计算机视觉系统在高级视觉问题上的推理能力。
The RAVEN dataset was developed by the Center for Vision, Cognition, Learning and Autonomy at the University of California, Los Angeles (UCLA), with the core goal of advancing machine intelligence by establishing connections between vision and structural, relational, and analogical reasoning. It consists of 70,000 Raven's Progressive Matrices problems, which are categorized into 7 distinct graphic configuration types. Each problem is annotated with 16 tree-structure annotations, leading to a total of 1,120,000 structural labels across the entire dataset. The dataset is specifically designed to test reasoning abilities rather than visual recognition: every image only contains simple grayscale objects, and reasoning rules are applied row by row, targeting the limitations of existing machines in short-term memory and compositional reasoning. By introducing structural representations, the RAVEN dataset offers a new paradigm for abstract reasoning for machines, with the aim of boosting the reasoning performance of computer vision systems when handling high-level visual problems.




