Mars-VL-Pairs
收藏资源简介:
该数据集是MarsRetrieval基准测试的任务1,旨在评估视觉语言模型(VLMs)在火星地理空间发现中的表现,专注于细粒度的图像-文本对齐。数据集包含2,287个配对的火星图像-文本样本,涵盖从全球轨道图像到漫游车级别图像的多尺度空间数据。每个样本包含图像URL、原始描述和精炼描述。该数据集适用于双向检索任务(文本→图像和图像→文本),并使用标准检索指标(如Recall@1、Recall@10、MRR和MedR)进行评估。数据集采用CC-BY-4.0许可,属于多模态和检索任务类别。
This dataset is Task 1 of the MarsRetrieval benchmark, which aims to evaluate the performance of Vision-Language Models (VLMs) in Martian geospatial discovery with a focus on fine-grained image-text alignment. The dataset comprises 2,287 paired Martian image-text samples, covering multi-scale spatial data spanning from global orbital images to rover-level images. Each sample contains an image URL, a raw description, and a refined description. This dataset supports bidirectional retrieval tasks (text→image and image→text) and is evaluated using standard retrieval metrics including Recall@1, Recall@10, MRR, and MedR. The dataset is licensed under CC-BY-4.0 and falls into the categories of multimodal and retrieval tasks.



