Perception and Grounding

DeblurGANv2

GAN-based blind motion deblurring for restoring sharper images from blurred inputs.

Tool Introduction

Core parameters, trigger timing, and visual before/after demo references.

Short Explanation

Input one blurred image and DeblurGANv2 generates a sharper restored image for downstream perception.

InputBlurred image

OutputDeblurred image

Trigger TimingTriggered on demand after the required input files and configuration are prepared.

RuntimePython / PyTorch

BeforeBlurred image

Prepare the scene, image, video, sensor stream, prompt, or configuration expected by the original project.

AfterDeblurred image

Read the produced visualization, prediction, map, trajectory, mask, grasp pose, or other documented artifact.

Preset Example

A quick-run style example for the documentation page.

Inputtools/deblurganv2/examples/blur.png

Promptmodel_name: fpn_inception

ExpectedA restored image file with reduced motion blur.

Parameters And Output

Readable controls and the meaning of each returned artifact.

Parameter Explanation

img_pathfile

Input blurred RGB image.

weights_pathpathtools/deblurganv2/weights/fpn_inception.h5

Checkpoint file used for inference.

model_nameselectfpn_inception

Generator architecture variant.

out_dirpath

Directory for restored images.

Output Explanation

deblurred_image

Restored image with sharper edges and textures.

output_path

Saved output file location.

How To Use

Official resources, deployment steps, academic context, citation, and source-reported benchmark numbers.

Resources

GitHubhttps://github.com/KupynOrest/DeblurGANv2 Code Downloadhttps://github.com/KupynOrest/DeblurGANv2/archive/refs/heads/master.zip arXivhttps://arxiv.org/abs/1908.03826 Pretrained Weightshttps://github.com/KupynOrest/DeblurGANv2#pre-trained-models

Deployment Notes

Clone the official repository and install Python dependencies from the README.
Download the official checkpoint and place it under tools/deblurganv2/weights/.
Run inference with repository-relative paths for image and checkpoint.
Store outputs under tools/deblurganv2/runs/ for catalog and evaluation.

Relative Path Example

python predict.py --img_path tools/deblurganv2/examples/blur.png --weights_path tools/deblurganv2/weights/fpn_inception.h5 --model_name fpn_inception --out_dir tools/deblurganv2/runs

Expected Result Shape

{
  "tool": "deblurganv2",
  "status": "ok",
  "results": [
    {
      "label": "Image deblurring",
      "score": 0.87,
      "output": "Deblurred image"
    }
  ],
  "timing": {
    "runtime": "The project emphasizes faster deblurring with FPN-based backbones, but the README benchmark table mainly exposes concrete GoPro quality numbers rather than a single universal latency figure.",
    "device": "documented in source benchmark when available"
  },
  "artifacts": {
    "visualization": "tools/deblurganv2/runs/visualization.png",
    "raw_predictions": "tools/deblurganv2/runs/predictions.json"
  }
}

Paper figure

Academic Info

Paper identity and contribution summary.

TitleDeblurGAN-v2: Deblurring (Orders-of-Magnitude) Faster and Better

AuthorsOrest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, Jiří Matas

VenueICCV 2019 / arXiv:1908.03826

ContributionIntroduces efficient generator backbones and improved adversarial training for practical blind motion deblurring.

Citation

@misc{deblurganv22019,
  title={DeblurGAN-v2: Deblurring (Orders-of-Magnitude) Faster and Better},
  author={Orest Kupyn and Volodymyr Budzan and Mykola Mykhailych and Dmytro Mishkin and Jiří Matas},
  year={2019},
  note={ICCV 2019 / arXiv:1908.03826},
  url={https://arxiv.org/abs/1908.03826}
}

Benchmark

Only compact, source-reported numbers are shown here.

Dataset	Metric	Value	Runtime	Source
GoPro test	PSNR / SSIM	29.55 / 0.934 for InceptionResNet-v2	fpn_inception.h5	Official pretrained-model table
GoPro test	PSNR / SSIM	28.17 / 0.925 for MobileNet; 28.03 / 0.922 for MobileNet-DSC	fpn_mobilenet.h5 and MobileNet-DSC	Official pretrained-model table

Artifacts

Official pretrained-model table, pretrained checkpoints, ICCV 2019 paper, and predict.py inference script.

Demo Images

Visual references from the original tool. Click any image to inspect the original size.