DIGR-Net

This is the code repository for DIGR-Net (Depth-induced Gap-reducing Network for RGB-D Salient Object Detection: An Interaction, Guidance and Refinement Approach)

Dataset Prapartion

create training dataset fold: open DIGR-Net Fold and create a new fold named by “dataset_dut”（datestet for training dataset(NUJ2000+NPLR+DUT-RGBD,2985 samples）, the structure of “dataset_dut” is: -- RGBD_for_test #test dataset ---NJU2K ----RGB ----depth ----GT ---NLPR ... -- RGBD_for train #training dataset ---RGB --- depth --- GT -- test_in_train #validation dataset ---NJU2K ----RGB ----depth ----GT ---NLPR ...

Train

###trianing with 2985 samples(NJU2K+NLPR+DUT-RGBD) With datstet for trainig preapread, open the fold "DIGR-Net" to find the file 'digr_train.py' , just run it!

###trianing with 2185 samples(NJU2K+NLPR+DUT-RGBD) open ./DIGR-Net/options.py and change the default fold from "./dataset_dut/RGBD_for_train/RGB/" to "./dataset/RGBD_for_train/RGB/",

and you should open digr_train.py to set

train_score=decode_csv('./score_folder/train_dut.csv') val_score=decode_csv('./score_folder/val_dut.csv')

to

train_score=decode_csv('./score_folder/train.csv') val_score=decode_csv('./score_folder/val.csv')

respectively.

Finally, you should set batch from 10 to 6, then you can run digr_train.py to train

datasets can be downloaded from BaiduNetDisk: datasets (yqut)

Test

open the fold "DIGR-Net" to find the file 'digr_test.py' （if you want to train with 2985 samples） of you can change the default test_path from "./dataset_dut/RGBD_for_test/" to "./dataset/RGBD_for_test/

Results

result maps can be downloaded from BaiduNetDisk: saliency maps (6odd)

Evaluation

find the evaluation method(matlab code) from evaluation code

Citation

@article{cheng2022ditr,
  title={Depth-induced Gap-reducing Network for RGB-D Salient Object Detection: An Interaction, Guidance and Refinement Approach},
  author={Cheng, Xiaolong and Zheng, Xuan and Pei, Jialun and Tang, He and Lyu, Zehua and Chen, Chuanbo},
  journal={IEEE Transactions on Multimedia},
  year={2022},
  publisher={IEEE}
}

Acknowledgement

We implement this project based on the code of 'Bbs-net: Rgb-d salient object detection with a bifurcated backbone strategy network', proposed by D.-P. Fan, Y. Zhai, A. Borji, J. Yang, and L. Shao in ECCV.