ProCrop 数据集
收藏资源简介:
ProCrop 数据集是微软研究院创建的一个大规模弱标注图像数据集,包含24.2万张图像。该数据集通过将专业图像外扩并迭代优化多种裁剪提案生成,旨在为图像裁剪任务提供丰富多样的高质量裁剪建议。数据集的创建过程涉及使用 GPT-4 生成文本描述和 SAM 生成多尺度构图掩码,然后输入到 ControlNet 进行图像外扩,最终生成多样化的裁剪提案。该数据集是迄今为止最大的公开可用的图像裁剪数据集,旨在解决当前图像裁剪数据集规模不足、多样性不足的问题,为图像美学和构图分析的研究提供有价值的资源。
The ProCrop dataset is a large-scale weakly-annotated image dataset created by Microsoft Research, containing 242,000 images. Generated by expanding professional images and iteratively optimizing multiple cropping proposals, this dataset aims to provide diverse, high-quality cropping suggestions for image cropping tasks. The dataset creation process involves using GPT-4 to generate text descriptions and SAM to generate multi-scale compositional masks, which are then fed into ControlNet for image expansion, ultimately yielding diverse cropping proposals. As the largest publicly available image cropping dataset to date, ProCrop addresses the limitations of existing image cropping datasets in terms of scale and diversity, serving as a valuable resource for research on image aesthetics and compositional analysis.

- 1ProCrop: Learning Aesthetic Image Cropping from Professional Compositions微软研究院 · 2025年



