UNICE Dataset Description

This is the dataset released with the paper: UNICE: Training A Universal Image Contrast Enhancer.

The UNICE dataset is crucial for training a universal and generalized model for various image contrast enhancement tasks, free of costly human labeling. It comprises HDR raw images used to render multi-exposure sequences (MES) and corresponding pseudo sRGB ground-truths via multi-exposure fusion.

Code: https://github.com/BeyondHeaven/UNICE

1. UNICEdataset.zip

• Type: Multi-Exposure Sequences (MES)
• Content: sRGB images rendered from HDR raw images using an emulated ISP pipeline.
• Structure: Each sequence contains multiple images of the same scene with varying exposure values (EVs), from -3EV to +3EV.
• Purpose: Serves as input data for training and evaluating exposure and contrast enhancement models.

2. pseudoGT.zip

• Type: Pseudo Ground Truths
• Content: High-quality sRGB images generated by fusing the MES using an ensemble of multi-exposure fusion (MEF) techniques.
• Purpose: Used as the target output (pseudo-GT) for supervised training of enhancement models.

3. pseudoGT_arniqa.csv

• Type: Pseudo Ground Truth Quality Scores
• Content: ARNIQA scores for each pseudoGT image, indicating perceptual quality.
• Purpose: Enables quality-aware selection of pseudoGTs. Low-quality samples (e.g., score < 0.5) can be filtered out to improve training.

Sample Usage

To download the dataset using Git LFS:

git lfs install
git clone https://huggingface.co/datasets/lahaina/UNICE

After downloading, you will find UNICEdataset.zip and pseudoGT.zip. For model training (e.g., as described in the associated code repository), you would typically extract these files and configure your dataset_folder to point to the extracted data. For instance, you might place the extracted contents into a directory like data/exposure and use it with the training scripts.