Exposure-slot: Exposure-centric representations learning with Slot-in-Slot Attention for Region-aware Exposure Correction (Official)

Donggoo Jung*, Daehyun Kim*, Guanghui Wang, Tae Hyun Kim†. (*Equal Contribution, †Corresponding author)

CVPR 2025

This repository contains the official PyTorch implementation of "Exposure-slot: Exposure-centric representations learning with Slot-in-Slot Attention for Region-aware Exposure Correction" accepted at CVPR 2025.

Exposure-slot is the first approach to leverage Slot Attention mechanism for optimized exposure-specific feature partitioning. We introduce the slot-in-slot attention that enables sophisticated feature partitioning and learning and exposure-aware prompts that enhance the exposure-centric characteristics of each image feature.

Our proposing method is the first approach to leverage Slot Attention mechanism for optimized exposure-specific feature partitioning. We introduce the slot-in-slot attention that enables sophisticated feature partitioning and learning and exposure-aware prompts that enhance the exposure-centric characteristics of each image feature. We provide validation code, training code, and pre-trained weights on three benchmark datasets (MSEC, SICE, LCDP).

Setting

Please follow these steps to set up the repository.

1. Clone the Repository

git clone https://github.com/kdhRick2222/Exposure-slot.git
cd Exposure-slot

2. Download Pre-trained models and Official Checkpoints

We utilize pre-trained models from Exposure-slot_ckpt.zip.

• Place the pre-trained models into the ckpt/ directory.

3. Prepare Data

For training and validating our model, we used SICE, MSEC, and LCDP dataset.

• SICE dataset

  We downloaded the SICE dataset from here.

  python prepare_SICE.py
  

  Make .Dataset_txt/SICE_Train.txt and .Dataset_txt/SICE_Test.txt for validation and training.

• MSEC dataset

  We downloaded the MSEC dataset from here.

  python prepare_MSEC.py
  

  Make .Dataset_txt/MSEC_Train.txt and .Dataset_txt/MSEC_Test.txt for validation and training.

• LCDP dataset

  We downloaded the LCDP dataset from here.

  python prepare_LCDP.py
  

  Make .Dataset_txt/LCDP_Train.txt and .Dataset_txt/LCDP_Test.txt for validation and training.

Inference and Evaluation

We provide 2-level and 3-level Exposure-slot model for each dataset (SICE, MSEC, LCDP).

python test.py --level=2 --dataset="MSEC"

Training

python train.py --gpu_num=0 --level=2 --dataset="MSEC"

Overall directory

├── ckpts
│ ├── LCDP_level2.pth
│ ├── LCDP_level3.pth
│ ├── MSEC_level2.pth
│ ├── MSEC_level3.pth
│ ├── SICE_level2.pth
│ └── SICE_level3.pth
│
├── config
│ ├── basic.py
│
├── data
│ ├── dataloaders.py
│ └── datasets.py
|
├── Dataset_txt
│ ├── LCDP_Train.txt
│ ├── LCDP_Test.txt
│ ├── MSEC_Train.txt
│ ├── MSEC_Test.txt
│ ├── SICE_Train.txt
│ └── SICE_Test.txt
|
├── utils
│ ├── scheduler_util.py
│ └── util.py
|
├── network_level2.py
├── network_level3.py
├── prepare_LCDP.py
├── prepare_MSEC.py
├── prepare_SICE.py
├── test.py
└── train.py

Citation

If you find our work useful in your research, please cite:

@inproceedings{jung2025Exposureslot,
  title={Exposure-slot: Exposure-centric representations learning with Slot-in-Slot Attention for Region-aware Exposure Correction},
  author={Donggoo Jung, Daehyun Kim, Guanghui Wang, Tae Hyun Kim},
  booktitle={Computer Vision and Pattern Recognition (CVPR)},
  year={2025}
}