Bongard-HOI
收藏资源简介:
Bongard-HOI数据集是由东北大学研究团队开发的一个新的视觉推理基准,专注于通过自然图像进行人类-物体交互(HOI)的组合学习。该数据集受到经典Bongard问题(BPs)的启发,具有两个显著特点:少样本概念学习和依赖上下文的推理。数据集精心策划了少样本实例,其中正负图像仅在动作标签上存在差异,使得仅通过物体类别识别无法完成基准测试。此外,Bongard-HOI设计了多个测试集,以系统地研究视觉学习模型的泛化能力,通过改变训练和测试集中HOI概念的重叠程度,从部分重叠到无重叠。Bongard-HOI为当前的视觉识别模型提出了重大挑战,旨在推动视觉推理研究,特别是在整体感知-推理系统和更好的表示学习方面。
The Bongard-HOI dataset is a novel visual reasoning benchmark developed by a research team from Northeastern University (China), focusing on compositional learning of human-object interaction (HOI) using natural images. Inspired by classic Bongard Problems (BPs), this dataset exhibits two prominent characteristics: few-shot concept learning and context-dependent reasoning. The dataset carefully curates few-shot instances, where positive and negative images differ only in their action labels, rendering the benchmark unsolvable solely through object category recognition. Furthermore, the Bongard-HOI dataset includes multiple test splits designed to systematically investigate the generalization performance of visual learning models, by adjusting the degree of overlap between HOI concepts in the training and test sets—ranging from partial overlap to zero overlap. Bongard-HOI poses significant challenges to contemporary visual recognition models, and aims to advance visual reasoning research, particularly in the areas of holistic perception-reasoning systems and enhanced representation learning.




