Awesome

ICNet-for-RGBD-SOD

This project provides the codes and results for 'ICNet: Information Conversion Network for RGB-D Based Salient Object Detection', TIP 2020. Paper link Homepage

Our code is implemented based on the Caffe of FlowNet2. You should first install and compile the caffe according to the FlowNet2.

Overview

Saliency maps and Measure results on 8 Datasets

We provide saliency maps (code: bqvj) and measure results (code: r0b6) of our ICNet on 5 datasets (STEREO, NJU2K, LFSD, DES and NLPR).

We also provide saliency maps & measure results (code: ujdp) of our ICNet on DUT-RGBD, SIP and SSD datasets.

You can use the evaluation tool to evaluate the result maps.

The parameter amount of our ICNet is 77.95M.

The FLOPs of our ICNet is 125.72G.

Testing

test_RGBD.prototxt/ is under models/.
Download the trained model (code: 6jz7) (RGBD_iter_25000.caffemodel), and put it under models/.
The datasets are under datasets/, we provide some testing examples on DES dataset.
Download depth2HHA.zip and unzip it, run depth2HHA.m to convert depth map to HHA.
Run test_matlab/test_ICNet.m.
Saliency maps are saved under salmaps/DES/.

Related works on RGB-D SOD

(TIP_2021_HAINet) Hierarchical Alternate Interaction Network for RGB-D Salient Object Detection.

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

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

Citation

    @ARTICLE{Li_2020_ICNet,
            author = {Li, Gongyang and Liu, Zhi and Ling, Haibin},
            title = {ICNet: Information Conversion Network for RGB-D Based Salient Object Detection},
            journal = {IEEE Transactions on Image Processing},
            year = {2020},
            volume = {29},
            pages = {4873-4884},}

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.