This project provides the code and results for 'Hierarchical Alternate Interaction Network for RGB-D Salient Object Detection', IEEE TIP 2021. Paper link Homepage

Network Architecture

Requirements

python2.7

pytorch 0.4.0

Our code is implemented based on the environment settings of CPD.

Usage

Modify the paths of VGG backbone (code: ego5) and datasets, then run train_HAI.py or test_HAI.py

Pre-trained model

Trained with NJU2K and NLPR (code: 4ntl)

Trained with NJU2K, NLPR and DUTLF-Depth (code: ae49)

RGB-D SOD Results Trained with NJU2K and NLPR

We provide results (code: a2as) of our HAINet on 5 datasets (STEREO1000, NJU2K, DES, NLPR and SIP) and additional 2 datasets (SSD and LFSD).

RGB-D SOD Results Trained with NJU2K, NLPR and DUTLF-Depth

We provide results (code: n35b) of our HAINet on 7 datasets (STEREO1000, NJU2K, DES, NLPR, SIP, DUTLF-Depth and ReDWeb-S).

RGB-T SOD Results

We apply our HAINet to RGB-T SOD, and provide results (code: s82s) of our HAINet on VT821 dataset trained with VT1000 dataset.

Evaluation Tool

You can use the evaluation tool to evaluate the above saliency maps.

Related works on RGB-D SOD

(ECCV_2020_CMWNet) Cross-Modal Weighting Network for RGB-D Salient Object Detection.

(TIP_2020_ICNet) ICNet: Information Conversion Network for RGB-D Based Salient Object Detection.

(Survey) RGB-D Salient Object Detection: A Survey.

Citation

    @ARTICLE{Li_2021_HAINet,
            author = {Gongyang Li and Zhi Liu and Minyu Chen and Zhen Bai and Weisi Lin and Haibin Ling},
            title = {Hierarchical Alternate Interaction Network for RGB-D Salient Object Detection},
            journal = {IEEE Transactions on Image Processing},
            year = {2021},
            volume = {30},
            pages = {3528-3542},}
            
            

If you encounter any problems with the code, want to report bugs, etc.

Please contact me at lllmiemie@163.com or ligongyang@shu.edu.cn.