Bongard-HOI

该数据集名为Bongard-HOI，是一个专注于从自然图像中学习人类与物体交互的复合视觉推理基准，旨在测试少样本学习和上下文依赖推理的能力。该基准包含了在训练期间完全未见过的动作和/或物体类别的具有挑战性的测试分割，这强调了需要对上下文依赖的推理和超越训练概念的泛化能力。该任务的目的是针对人类与物体交互进行少样本视觉推理。

The dataset named Bongard-HOI is a compositional visual reasoning benchmark dedicated to learning human-object interactions (HOI) from natural images, designed to test few-shot learning and context-dependent reasoning abilities. This benchmark includes challenging test splits that feature action and/or object categories entirely unseen during training, which underscores the need for context-dependent reasoning and generalization beyond the concepts seen during training. The core objective of this task is to conduct few-shot visual reasoning focused on human-object interactions.