PointPrompt
收藏资源简介:
PointPrompt是由乔治亚理工学院创建的一个视觉提示数据集,涵盖了多个图像类别和多个标注者。该数据集包含6000张图像,来自COCO、NDD20、医学和地震等多个公开数据库。数据集的创建过程包括使用SAM模型进行标注,生成包含提示点坐标、掩码和IoU分数的序列。PointPrompt旨在解决视觉提示策略的有效性问题,特别是在自动化方法与人类标注之间的比较,以及通过微调提高SAM模型性能的研究。
PointPrompt is a visual prompt dataset created by the Georgia Institute of Technology, covering multiple image categories and involving multiple annotators. This dataset contains 6000 images sourced from multiple public databases including COCO, NDD20, medical imaging datasets, and seismology-related datasets. The dataset construction process uses the SAM model for annotation, generating sequences that include prompt point coordinates, masks, and IoU scores. PointPrompt aims to address the effectiveness of visual prompt strategies, particularly comparative studies between automated annotation methods and human annotations, as well as research on improving SAM model performance via fine-tuning.

- 1Benchmarking Human and Automated Prompting in the Segment Anything Model乔治亚理工学院 · 2024年



