Skip to content

This research is based on efficient saliency detection using deformable convolutions.

License

Notifications You must be signed in to change notification settings

tanveer-hussain/EfficientSOD

Repository files navigation

Densely Deformable Efficient Salient Object Detection Network

PWC

Contents

  1. Introduction
  2. The proposed DDNet
  3. Requirements and how to run?
  4. Quantitative and qualitative comparisons
  5. Citation and acknowledgements

Introduction

In this paper, inspired by the best background/foreground separation abilities of deformable convolutions, we employ them in our Densely Deformable Network (DDNet) to achieve efficient SOD. The salient regions from densely deformable convolutions are further refined using transposed convolutions to optimally generate the saliency maps. Quantitative and qualitative evaluations using the recent SOD dataset against 22 competing techniques show our method’s efficiency and effectiveness.

The proposed DDNet

The proposed DDNet uses three main blocks to generate optimal saliency. Firstly, two dense convolution blocks represent lowlevel features of the input RGB images. Then we propose densely connected deformable convolutions to learn effective features of salient regions and their corresponding boundaries. Finally, we employ transpose convolution and upsampling to generate the resultant saliency image, refer to the figure below:

The proposed DDNet schematic diagram

Requirements and how to run?

For the trained models and high-resolution images, please visit: https://drive.google.com/drive/folders/1aigSE0nLKfYlAbl9CIk4mSpCxneS2fUw?usp=sharing

Make a folder TrainedModels in the same repository and download the pretrained DDNet weights and model from the above link^.

Quantative and Visual comparisons

Comparison with SOTA on SIP dataset Visual results on challenging SIP test images

Citation and acknowledgements

@misc{hussain2021densely,
      title={Densely Deformable Efficient Salient Object Detection Network}, 
      author={Tanveer Hussain and Saeed Anwar and Amin Ullah and Khan Muhammad and Sung Wook Baik},
      year={2021},
      eprint={2102.06407},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Releases

No releases published

Packages

No packages published