OCORD
收藏资源简介:
OCORD数据集是由北京邮电大学创建的高分辨率真实世界数据集,专为对象移除任务设计。该数据集通过固定相机角度的长时视频捕捉构建,包含图像、背景和掩码对,数据来源于真实世界的自然场景,准确反映了现实世界中的物理变化。数据集采用自动化标注工具(如Grounding-DINO、SAM和MASA)进行高效标注,显著减少了人工成本。OCORD数据集旨在解决现有合成数据集在真实场景中的不一致性问题,提升对象移除模型在复杂自然场景中的性能,广泛应用于图像修复和生成领域。
The OCORD dataset is a high-resolution real-world dataset developed by Beijing University of Posts and Telecommunications (BUPT), specifically tailored for object removal tasks. It is constructed through long-duration video capture with a fixed camera angle, and includes image-background-mask triplets sourced from real-world natural scenes, accurately reflecting physical changes in actual environments. The dataset utilizes automated annotation tools including Grounding-DINO, SAM and MASA for efficient labeling, which greatly reduces manual labor costs. The OCORD dataset aims to solve the inconsistency problem of existing synthetic datasets in real-world scenarios, improve the performance of object removal models in complex natural scenes, and is widely applied in the fields of image inpainting and generation.




