TDBench

该数据集名为TDBench，是一个专注于自上而下图像理解的视觉-语言模型（VLMs）的全面基准测试，它由公共自上而下视角数据集和高质量模拟图像构建而成，包含了跨越十个评估维度的视觉问答对。TDBench还包括一种名为RotationalEval（RE）的新评估策略，用于评估模型在不同图像方向下的性能，通过旋转不变性确保了严格的评估。这一全面基准测试具有多个评估维度，其任务涵盖了自上而下图像理解和视觉问答。

This dataset, named TDBench, is a comprehensive benchmark for Vision-Language Models (VLMs) focused on top-down image understanding. It is constructed from public top-down view datasets and high-quality simulated images, and contains visual question-answer pairs spanning ten evaluation dimensions. TDBench also includes a novel evaluation strategy named RotationalEval (RE) for evaluating model performance under different image orientations, which ensures rigorous evaluation via rotational invariance. This comprehensive benchmark has multiple evaluation dimensions, with its tasks covering top-down image understanding and visual question answering.