bcmi/Image-Harmonization-Dataset-iHarmony4

[CVPR 2020] The first large-scale public benchmark dataset for image harmonization. The code used in our paper "DoveNet: Deep Image Harmonization via Domain Verification", CVPR2020. Useful for image harmonization, image composition, etc.

What it solves

This project provides iHarmony4, the first large-scale dataset specifically designed for image harmonization. Image harmonization is the process of adjusting the appearance of a foreground object in a composite image so that it looks consistent and natural with the background region, solving the problem of high-quality training data being too time-consuming and skill-demanding to produce manually.

How it works

The project releases a collection of four sub-datasets (HCOCO, HAdobe5k, HFlickr, and Hday2night) containing synthesized composite images, foreground masks, and their corresponding real images. To create these, the authors used various color transfer methods—including global color transfer in decorrelated and RGB color spaces, cumulative histogram mapping, and IDT regrain color transfer—to synthesize realistic composite images from existing datasets like Microsoft COCO and MIT-Adobe FiveK.

Who it’s for

This is primarily for researchers and developers working on computer vision and image editing, specifically those developing models for image composition, object insertion, and image harmonization.

Highlights

  • Large-scale data: Includes over 65,000 training images across four diverse sub-datasets.
  • Diverse sources: Combines data from COCO, Adobe FiveK, Flickr, and day-to-night transitions.
  • Integrated Tooling: Provides a PyTorch implementation of DoveNet, a deep image harmonization model, along with pre-trained weights.
  • Baseline Comparisons: Includes code and implementation details for several baseline models (e.g., DIH, S2AM, SSH) to facilitate fair benchmarking.

Related

  • Project
    bcmi/libcom

    An image composition toolbox that provides a unified set of tools for realistic object insertion, including harmonization, shadow generation, and placement assessment.

  • Project
    Linfeng-Tang/Image-Fusion

    A comprehensive collection of research papers and code implementations for image fusion, covering multi-modal, digital photography, and remote sensing image combination techniques.

  • Project
    hanna-xu/U2Fusion

    A unified unsupervised image fusion network that merges multi-modal, multi-exposure, and multi-focus images into a single high-quality image without requiring labeled training data.

  • Dispatch
    Image Similarity with Hugging Face Datasets and Transformers

    Hugging Face demonstrates how to build an image similarity system using the Transformers and Datasets libraries by computing dense vector embeddings and measuring cosine similarity.

  • Project
    XiandaGuo/OpenStereo

    A comprehensive benchmark and framework for stereo matching that integrates 17 datasets and multiple state-of-the-art models for depth estimation.