DEROBA is a large-scale real-world dataset for reflection generation, i.e., generating plausible reflection for the inserted foreground object, which is particularly valuable for image composition (object insertion). DEROBA contains 16,791 different images and 21,016 object-reflection pairs. The figure below shows several examples. From left to right in each example, we show the composite image, the foreground mask, the reflection mask, and the ground-truth image.
We collect original_image with object reflections from pixabay and annotate the foreground_mask, reflection_mask. Then, for each object-reflection pair, we employ image inpainting model to erase the object and reflection, resulting in inpainted_image. Because inpainting causes color disturbation, we apply image inpainting model with empty mask to get the ground-truth_image. We crop the foreground from ground-truth_image and paste it on the inpainted_image to obtain composite_image.
We provide two versions: the full-resolution version and the 512-resolution version. The full-resolution version is available on: [Baidu_Cloud] (access code: bcmi) or [Dropbox]. The 512-resolution version is available on: [Baidu_Cloud] (access code: bcmi) or [Dropbox]. We also provide the training-test split. Each version has the following file structure:
├── composite_image:
├── alpacas-7604526_box0.png
├── alpacas-7604526_box1.png
├── ……
├── foreground_mask:
├── alpacas-7604526_box0.png
├── alpacas-7604526_box1.png
├── ……
├── reflection_mask:
├── alpacas-7604526_box0.png
├── alpacas-7604526_box1.png
├── ……
├── ground-truth_image:
├── alpacas-7604526_box0.png
├── alpacas-7604526_box1.png
├── ……
├── inpainted_image:
├── alpacas-7604526_box0.png
├── alpacas-7604526_box1.png
├── ……
├── original_image:
├── alpacas-7604526_box0.png
├── alpacas-7604526_box1.png
├── ……
├── train.txt
└── test.txt
If you use our DEROBA dataset, please cite the following BibTeX [arxiv]:
@article{niu2021making,
title={Making images real again: A comprehensive survey on deep image composition},
author={Niu, Li and Cong, Wenyan and Liu, Liu and Hong, Yan and Zhang, Bo and Liang, Jing and Zhang, Liqing},
journal={arXiv preprint arXiv:2106.14490},
year={2021}
}