X-DECODE: EXtreme Deblurring with Curriculum Optimization and Domain Equalization

Evaluation - Train/Test on Extreme-GOPRO::

BL	Metric	DeblurGAN	DeblurGANv2	NAFNet	Ours
19	SSIM	0.282	0.668	0.642	0.764
	PSNR	19.51	23.94	22.99	24.90
25	SSIM	0.258	0.635	0.577	0.674
	PSNR	19.09	22.88	21.50	23.64
29	SSIM	0.240	0.618	0.545	0.613
	PSNR	18.84	22.33	20.76	22.68

Evaluation - Train on EXTREME-GOPRO and Test on EXTREME-KITTI:

Method	SSIM	PSNR
DeblurGAN [5]	0.256	16.24
DeblurGANv2 [6]	0.466	18.45
NAFNet [8]	0.428	18.18
Ours (Without CL)	0.386	18.15
Ours (With CL)	0.549	20.71

Training and Testing on Extreme GoPRO Dataset

First install all requirements:

pip install -r requirements.txt

To train the model on the Extreme GoPRO dataset, use the following command:

python TrainTestSameDomainGoPro/main.py --dataset="<dataset_name>"

To test the model:

python TrainTestSameDomainGoPro/test.py

Training on Extreme Cityscapes and PascalVOC, Testing on KITTI Dataset

• Supported dataset options: Cityscapes, GoPRO, PascalVOC.
• Use this to train the model using different curriculum learning strategies:

python TrainTestOnDiffDomain/main.py --dataset="dataset_dir/" --curr_lear="<curriculum_type>"

Supported values for --curr_lear:

• linear : Linear curriculum learning
• stepwise : Step-wise curriculum learning
• slow-stepwise : Slower step-wise curriculum learning
• expo : Exponential curriculum learning
• sigmoid : Sigmoid-based curriculum learning
• none : Training without curriculum learning

• Supported datasets: Cityscapes, GoPRO, PascalVOC.

To test the model:

python TrainTestOnDiffDomain/test_new_range_of_blur.py --dataset_dir="dataset_dir/" --model_dir="model_dir/"

It also generates a CSV file containing the computed SSIM and PSNR for each image pair and the final mean SSIM and PNSR.

Dataset

The following datasets were used for training and evaluation:

Dataset	Purpose	Description
GoPRO	Primary training & evaluation	Each blurred image has a corresponding sharp image. Extreme motion blur levels (BL19 to BL29) are simulated using the Albumentations library.
PascalVOC-2012	Cross-domain training	Standard object recognition dataset adapted for deblurring experiments.
Cityscapes	Cross-domain training	Urban street scene dataset adapted for deblurring tasks.
KITTI	Cross-domain testing	Major test dataset for evaluating performance across domains.

Sample Images

Extreme-GoPRO - Samples (First column sharp image and rest others is obtained after applying Blur Level 19, 25 and 29):

KITTI - Samples (First column sharp image and rest others is obtained after applying Blur Level 19, 25 and 29):

Dataset Download Links

Dataset	Download Link
Extrene GoPRO	Download
Extreme PascalVOC-2012	Download
Extreme Cityscapes	Download
Extreme KITTI	Download

Pretrained Models

Pretrained models for different datasets:

Trained on	Learning type	Download Link
GoPRO extreme Blurred	Step-wise (Train/Test on GoPRO)	Download
GoPRO extreme Blurred	Linear Curriculum Learning (Test on KITTI)	Download
Cityscapes extreme Blurred	Linear Curriculum Learning (for KITTI testing)	Download
Pascal VOC extreme Blurred	Linear Curriculum Learning (for KITTI testing)	Download

Pretrained models for different curriculum learning. All Tested on Kitti Dataset:

Model Description	Curriculum Learning	Download Link
Cityscapes Extreme Blurred	Step-wise	Download
Cityscapes Extreme Blurred	Slower Step-wise	Download
Cityscapes Extreme Blurred	Linear	Download
Cityscapes Extreme Blurred	Exponential	Download
Cityscapes Extreme Blurred	Sigmoid	Download

Comparison of Curriculum Learning Techniques on Extreme-KITTI Dataset, Trained on Extrem-Cityscapes Dataset Results

Curriculum Learning Method	SSIM	PSNR
Step-wise	0.619	21.97
Slower Step-wise	0.468	18.91
Linear	0.628	22.20
Sigmoid	0.405	18.02
Exponential	0.614	21.79

Citation

If you use the X-DECODE data or code please cite:

@article{xdecode2025,
  title = {X-DECODE: EXtreme Deblurring with Curriculum Optimization and Domain Equalization},
  author = {Gautam Sushant and Chen Jingdao},
  journal = {ArXiv e-prints},
  eprint = {2504.08072},
  year = {2025},
  url={https://arxiv.org/abs/2504.08072}
}