This repository provides the iHarmony4 dataset, which is a large-scale dataset designed for image harmonization tasks. Image harmonization involves adjusting the appearance of a foreground in a composite image so that it is consistent with the background (in color, tone, illumination, etc.). The iHarmony4 dataset comprises four sub-datasets (HCOCO, HAdobe5k, HFlickr, Hday2night), each making composite images by combining a foreground from one image with a background from another, along with associated ground truth harmonized images and foreground masks. The dataset is intended as a benchmark resource to enable and standardize research in image harmonization. Each composite sample has: composite image, foreground mask, and corresponding real harmonized image.
Features
- Includes 4 sub-datasets: HCOCO, HAdobe5k, HFlickr, Hday2night
- Each composite sample has: composite image, foreground mask, and corresponding real harmonized image
- Large scale: ~65,742 training samples total, ~7,404 test samples across sub-datasets
- Augmentation via SycoNet (to synthesize more composites) discussed in repo / readme
- Supports color variation / illumination changes across composites
- Provides dataset download links (e.g. Baidu Cloud, Dropbox)