hikkilover/OpenImage-FCO

---
license: apache-2.0
task_categories:
- image-to-image
language:
- en
size_categories:
- 10M<n<100M
---

The Dataset of ICCV2025 paper "Toward Better Out-painting: Improving the Image Composition with Initialization Policy Model".

This dataset is specifically designed for the Foreground-Conditioned Outpainting (FCO) task. It contains over 580,000 image-text pairs and more than 800,000 foreground object masks. We leverage BLIP2 and RAM for image captioning and tagging, while GroundingDINO and HQ-SAM are employed to generate high-precision object masks.

All images are sourced from OpenImage-v7. To enhance data efficiency, we have filtered out samples that are clearly unsuitable for the FCO task—such as masks that are excessively large or small, and masks covering non-object regions (e.g., sky and ground). Additionally, we have merged regions with identical semantics within the same image.

P.S. Since our research primarily serves the background generation of commodities, nearly all masks related to humans have also been filtered out.

许可证：Apache-2.0
任务类别：
- 图像到图像（image-to-image）
语言：
- 英语
样本规模：
- 10M<n<100M

本数据集为ICCV2025论文《面向更优图像外扩：基于初始化策略模型优化图像合成》的配套数据集。
该数据集专为前景条件图像外扩（Foreground-Conditioned Outpainting, FCO）任务设计，包含超58万组图像-文本对与80余万张前景目标掩码。我们借助BLIP2与RAM完成图像字幕生成与标签标注，并采用GroundingDINO与HQ-SAM生成高精度目标掩码。
所有图像均源自OpenImage-v7。为提升数据利用效率，我们过滤掉了明显不适用于FCO任务的样本——例如尺寸过大或过小的掩码、覆盖非目标区域（如天空、地面）的掩码。此外，我们还合并了单张图像内语义一致的区域。
备注：由于本研究主要服务于商品背景生成任务，几乎所有与人类相关的掩码均已被过滤。