FineCops-Ref
收藏资源简介:
FineCops-Ref是由电子科技大学创建的一个新的细粒度组合参考表达理解数据集。该数据集包含9605条正样本和9814条负样本表达,以及8507张负样本图像。数据集的设计旨在测试多模态大语言模型在对象类别、属性和多跳关系上的细粒度推理能力,并通过负样本测试模型在目标对象不在图像中的情况下正确拒绝的能力。数据集的创建过程包括路径生成、表达生成和负样本生成,通过精细的编辑和生成技术确保数据的高质量。该数据集主要应用于视觉推理和跨模态交互策略的开发,旨在提升多模态大语言模型的视觉接地能力。
FineCops-Ref is a novel fine-grained compositional referring expression understanding dataset created by the University of Electronic Science and Technology of China. This dataset comprises 9605 positive referring expressions, 9814 negative referring expressions, and 8507 negative sample images. It is designed to evaluate the fine-grained reasoning capabilities of multimodal large language models (LLMs) on object categories, attributes, and multi-hop relational reasoning, as well as to test the model's ability to correctly reject scenarios where the target object is absent from the image via negative samples. The dataset construction process includes path generation, expression generation and negative sample generation, with high data quality ensured through meticulous editing and generative techniques. This dataset is primarily utilized for the development of visual reasoning and cross-modal interaction strategies, aiming to enhance the visual grounding capabilities of multimodal LLMs.

- 1FineCops-Ref: A new Dataset and Task for Fine-Grained Compositional Referring Expression Comprehension电子科技大学 · 2024年



