Xinyueliii/HADataset

---
readme: Rethinking Artifact Mitigation in HDR Reconstruction
license: mit
task_categories:
- image-to-image
- object-detection
- image-segmentation
language:
- en
tags:
- hdr
- artifact-detection
- image-restoration
- high-dynamic-range
pretty_name: HADataset
size_categories:
- 100B<n<1T
source_datasets:
- original
configs:
- config_name: default
data_files:
- split: train
path: data/Training/**
- split: test
path: data/Test/**
---

# Rethinking Artifact Mitigation in HDR Reconstruction: From Detection to Optimization
#### IEEE Transactions on Image Processing (TIP), 2025

[](https://ieeexplore.ieee.org/document/11301923)
[](https://github.com/xinyueliii/hdr-artifact-detect-optimize)
## Dataset Description

**HADataset** is the first dedicated High Dynamic Range (HDR) artifact dataset designed to address the challenge of ghosting artifacts in HDR reconstruction. It explicitly provides per-pixel artifact annotations, enabling the development of detection-aware optimization strategies.

This dataset was introduced in the paper "Rethinking Artifact Mitigation in HDR Reconstruction: From Detection to Optimization". It serves two main purposes:
1. **Artifact Detection:** Training models (like HADetector) to localize artifacts.
2. **HDR Reconstruction:** providing diverse multi-exposure Low Dynamic Range (LDR) image sets for testing and training reconstruction algorithms.

### Key Features
* **Total LDR Sets:** 1,213 diverse multi-exposure sets.
* **Annotated Pairs:** 1,765 HDR image pairs with per-pixel artifact annotations.
* **Diverse Sources:** Includes artifacts from Kalantari’s dataset, our own collected scenes, and Tel’s dataset.
* **Model-Agnostic:** Includes artifacts generated by various state-of-the-art models (AHDR, CA-ViT, SCTNet).

## Dataset Structure

The HADataset consists of two main components:

### 1. HADataset-LDRsets (Source LDR Images sets)
This component includes 1,216 LDR sets captured for HDR inference.
* **Training Set:** 970 sets
* **Test Set:** 243 sets

Each set typically contains 3 exposure brackets (short, medium, long) in `.tif` format along with an `exposure.txt` file.

### 2. HADataset-HDRArtifactDetection (HDR images and Annotations)
This component is designed for the artifact detection task. It contains ground truth (GT) artifact maps and the corresponding HDR images (Tp). It is categorized into two perspectives:

#### Content Perspective (3 Subsets)
Based on the origin of the scene:
* `HADataset-content-Kal`: Scenes from Kalantari's dataset.
* `HADataset-content-Ours`: Scenes collected by the authors.
* `HADataset-content-Tel`: Scenes from Tel's dataset.

#### Model Perspective (3 Subsets)
Based on the model that generated the artifacts:
* `HADataset-content-AHDR`
* `HADataset-content-CaViT`
* `HADataset-content-SCTNet`

## Citation

If you use this dataset in your research, please cite our paper:

```bibtex
@ARTICLE{11301923,
author={Li, Xinyue and Ni, Zhangkai and Wu, Hang and Yang, Wenhan and Wang, Hanli and He, Lianghua and Kwong, Sam},
journal={IEEE Transactions on Image Processing},
title={Rethinking Artifact Mitigation in HDR Reconstruction: From Detection to Optimization},
year={2025},
volume={34},
pages={8435-8446},
doi={10.1109/TIP.2025.3642557}
}