MMPerspective
收藏资源简介:
MMPerspective是一个专为评估多模态大型语言模型(MLLMs)对透视理解能力而设计的基准数据集。该数据集由2711个真实世界和合成图像实例组成,包含5083个问题-答案对,旨在探索关键能力,如灭点感知和计数、透视类型推理、三维空间中的线关系理解等。通过全面评估43个最先进的MLLMs,揭示了模型在表面感知任务上的能力,以及在组合推理和保持空间一致性方面的局限性。MMPerspective为诊断和推进视觉语言系统中的空间理解提供了一个宝贵的测试平台。
MMPerspective is a benchmark dataset specifically designed for evaluating the perspective understanding capabilities of multimodal large language models (MLLMs). This dataset consists of 2711 real-world and synthetic image instances, along with 5083 question-answer pairs, and aims to explore key capabilities such as vanishing point perception and counting, perspective type reasoning, and understanding of line relationships in 3D space. Through a comprehensive evaluation of 43 state-of-the-art MLLMs, this work reveals the models' strengths on surface-level perception tasks, as well as their limitations in compositional reasoning and maintaining spatial consistency. MMPerspective serves as a valuable testbed for diagnosing and advancing spatial understanding in vision-language systems.

- 1MMPerspective: Do MLLMs Understand Perspective? A Comprehensive Benchmark for Perspective Perception, Reasoning, and Robustness罗切斯特大学 · 2025年



