grow-ai-like-a-child/perceptual-constancy

感知恒常性（Perceptual Constancy）是一个多模态基准数据集，旨在评估大型视觉语言模型（VLMs）的高级感知不变性。它通过变化的感官外观来探测模型对物理和几何稳定性的理解。该数据集包含253个样本，有两种模态：图像和短视频剪辑，以及两种问题格式：选择题和判断题。每个问题都测试模型是否能够在诸如视点、颜色、方向、大小或遮挡等变换下泛化一致的属性。

The Perceptual Constancy is a multimodal benchmark designed to evaluate high-level perceptual invariance in large vision-language models (VLMs). It probes a models understanding of physical and geometric stability under varying sensory appearances. The dataset consists of 253 samples with two modalities: images and short video clips, and two question formats: multiple-choice and true/false. Each question tests whether the model can generalize consistent properties across transformations such as viewpoint, color, orientation, size, or occlusion.