IDMR-bench
收藏arXiv2025-09-30 收录
下载链接:
https://github.com/BwLiu01/IDMR
下载链接
链接失效反馈官方服务:
资源简介:
该数据集是为了对实例驱动的多模态图像检索(IDMR)进行基准测试而开发的,它使用了真实世界的物体追踪和第一人称视角的视频数据。该数据集包含了丰富多样的上下文信息,以支持细粒度的实例级一致性,并采用跨领域合成方法构建。规模上,该数据集包含了120万交织的文本-图像对,其中557,000条为合成数据,662,000条来自MMEB。其任务定位于实例驱动的多模态图像检索。
This dataset is developed for benchmarking instance-driven multimodal image retrieval (IDMR), leveraging real-world object tracking and first-person perspective video data. Constructed via cross-domain synthesis approaches, it encompasses rich and diverse contextual information to enable fine-grained instance-level consistency. In terms of scale, the dataset contains 1.2 million interleaved text-image pairs, among which 557,000 are synthetic samples and 662,000 are derived from MMEB. The task of this benchmark is focused on instance-driven multimodal image retrieval.
提供机构:
The authors of the paper



