Skip to content

Latest commit

 

History

History
67 lines (41 loc) · 1.93 KB

README.md

File metadata and controls

67 lines (41 loc) · 1.93 KB
Raw

MPI: Multi-receptive and Parallel Integration for Salient Object Detection

Code for paper "MPI: Multi-receptive and Parallel Integration for Salient Object Detection", by Han Sun, Jun Cen , Ningzhong Liu, Dong Liang, and Huiyu Zhou.

Requirement

  • python-3.5
  • pytorch-1.4.0
  • torchvision
  • numpy
  • apex
  • cv2

Usage

  • Clone this repo into your workstation
git clone https://github.com/NuaaCJ/MPI.git

  • Set your dataset format as follows:

    ECSSD\

    --Image\ *.jpg

    --Mask\ *.png

    DUTS\

    --Image\ *.jpg

    --Mask\ *.png

    ...

- training

  1. Download the pre-trained model for resnet50 resnet50-19c8e357.pth (passwd: resi)

  2. Generate edge maps for the training set, or download the file we provide DUT_TR_edges (passwd: edge)

  3. Modify MPI\train_mpi.py to change both the dataset path and the file save path to your own real path

  4. run train_mpi.py

python3.5 train_mpi.py
- test

  1. Download our trained model MPI_model (passwd: mpim) and put it into folder MPI\models

  2. Modify the dataset path and file save path in the MPI\test.py and MPI\main_function.m to your own real paths

  3. run test.py, then the saliency maps will be generated under the corresponding path, and the evaluation scores for the model on the test dataset will be stored in result.txt

python3.5 test.py

The result saliency maps

Here are saliency maps of our model on five different datasets (DUTS, ECSSD, DUT-OMRON, HKU-IS, PASCAL-S) The result saliency maps (passwd: maps)

Acknowledgements

The evaluation codes (MPI\*.m) we used are provided by F3net