RGB-Th-Bench
收藏资源简介:
RGB-Th-Bench是一个针对视觉-热理解任务的高质量密集型多模态基准,包含58个RGB-热图像样本对和1624个问题。这些问题覆盖了14个核心评估或技能维度,每个样本对针对每个技能维度有四个独立的是非问题。数据集的问题和答案都是通过领域专家监督下手动注释的,旨在评估视觉语言模型在不依赖热输入和依赖热输入情况下的性能,涵盖场景理解、详细物体存在、实例属性等维度。
RGB-Th-Bench is a high-quality dense multimodal benchmark for visual-thermal understanding tasks. It contains 58 RGB-thermal image pairs and 1,624 questions. These questions cover 14 core evaluation or skill dimensions, and each image pair has four independent true-false questions for each skill dimension. Both the questions and answers are manually annotated under the supervision of domain experts, aiming to evaluate the performance of vision-language models both with and without thermal input, covering dimensions such as scene understanding, detailed object existence, and instance attributes.

- 1RGB-Th-Bench: A Dense benchmark for Visual-Thermal Understanding of Vision Language ModelsAalto University, KTH Royal Institute of Technology, Detectium Oy · 2025年



