hikkilover/OpenImage-FCO

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.