Deblur-YOLO

This is the official Pytorch implementation for our paper:

"Deblur-YOLO: Real-Time Object Detection with Efficient Blind Motion Deblurring" [IJCNN 2021 Paper Link]

Abstract

Object detection has been a traditional yet open computer vision research field. In intensive studies, object detection models have achieved promising results regarding recognition accuracy and inference speed. However, previous state-of-the-art algorithms fail to operate at blurry images. In this work, we propose Deblur-YOLO, an efficient, YOLO-based and detection-driven approach robust to motion blur photographs. We introduce a generative adversarial network with a dilated feature pyramid generator, a pair of multi-scale discriminators with spectral normalization, and a detection discriminator. We design a new image quality metric called Smooth Peak Signal-to-Noise Ratio (SPSNR) for measuring the smoothness of the reconstructed image. Empirical studies on benchmark datasets demonstrate Deblur-YOLO's superiority. On COCO 2014, Set 5 and Setl4, Deblur-YOLO achieves leading results for parameters, deblurring time, PSNR, SPSNR and SSIM. We also visually display the excellence of our deblurring performance to competing models.

Visual Comparisons

Toy Visual Comparison

From left to right: the object detection result on blurry image, blurry image restored by Deblur-YOLO, and the groundtruth image.

Visual Comparison on Blurred COCO 2014

From top left to bottom right: the object detection result on Clean Image, Blurred Image, DeepDeblur, SRN Deblur, DynamicDeblur, DeblurGANv2(I-R), DeblurGANv2(M), Deblur-YOLO

Model Architecture

Click the links below to see the model architecture

Prerequisite

• Windows or Linux
• CUDA 10.0+
• Python 3
• Pytorch 1.0+
• torchvision
• torchsummary
• opencv-python
• numpy
• albumentations
• scikit-image
• glog
• fire

Dataset

Training Dataset

Testing Dataset

Training

Testing

Hyperparameters

Citation

@inproceedings{zheng2021deblur,
  title={Deblur-YOLO: Real-Time Object Detection with Efficient Blind Motion Deblurring},
  author={Zheng, Shen and Wu, Yuxiong and Jiang, Shiyu and Lu, Changjie and Gupta, Gaurav},
  booktitle={2021 International Joint Conference on Neural Networks (IJCNN)},
  pages={1--8},
  year={2021},
  organization={IEEE}
}

TODO List

• Add Dependencies
• Upload Model Architecture Figure
• Upload Visual Comparisons
• List important hyperparameters
• Upload Training Dataset
• Upload Testing Dataset
• Update Code
• Testing on Video
• Upload Pretrained Weight
• Finalize readme

Acknowledgement

This code is heavily based upon DeblurGANv2. We thanks the authors for sharing their code.