Awesome

BBS-Net

BBS-Net: RGB-D Salient Object Detection with a Bifurcated Backbone Strategy Network

<img src="Images/pipeline.png" width="80%"/> Figure 1: Pipeline of the BBS-Net.

1. Requirements

Python 3.7, Pytorch 0.4.0+, Cuda 10.0, TensorboardX 2.0, opencv-python

2. Data Preparation

Download the raw data from Baidu Pan [code: yiy1] or Google Drive and trained model (BBSNet.pth) from Here [code: dwcp]. Then put them under the following directory:
```
 -BBS_dataset\ 
   -RGBD_for_train\  
   -RGBD_for_test\
   -test_in_train\
 -BBSNet
   -models\
   -model_pths\
      -BBSNet.pth
   ...
   
```
Note that the depth maps of the raw data above are not normalized. If you train and test using the normalized depth maps, the performance will be improved.

3. Training & Testing

Train the BBSNet:

python BBSNet_train.py --batchsize 10 --gpu_id 0
Test the BBSNet:

python BBSNet_test.py --gpu_id 0

The test maps will be saved to './test_maps/'.
Evaluate the result maps:

You can evaluate the result maps using the tool in Python_GPU Version or Matlab Version.
If you need the codes using VGG16 and VGG19 backbones, please send to the email (zhaiyingjier@163.com). Please provide your Name & Institution. Please note the code can be only used for research purpose.

4. Results

4.1 Qualitative Comparison

<img src="Images/resultmap.png" width="80%"/> Figure 2: Qualitative visual comparison of the proposed model versus 8 SOTA models. <img src="./Images/detailed-comparisons.png" width="80%"/> Table 1: Quantitative comparison of models using S-measure max F-measure, max E-measureand MAE scores on 7 datasets.

4.2 Results of multiple backbones

<img src="./Images/backbone_result.png" width="80%"/> Table 2: Performance comparison using different backbones.

4.3 Download

Test maps of the above datasets (ResNet50 backbone) can be download from here [code: qgai ].
Test maps of vgg16 and vgg19 backbones of our model can be download from here [code: zuds ].
Test maps of DUT-RGBD dataset (using the proposed training-test splits of DMRA) can be downloaded from here [code: 3nme ].

5. Citation

Please cite the following paper if you use this repository in your reseach.

@inproceedings{fan2020bbsnet,
title={BBS-Net: RGB-D Salient Object Detection with a Bifurcated Backbone Strategy Network},
author={Fan, Deng-Ping and Zhai, Yingjie and Borji, Ali and Yang, Jufeng and Shao, Ling},
booktitle={ECCV},
year={2020}
}

For more information about BBS-Net, please read the Manuscript (PDF) (Chinese version[code:0r4a]).
Note that there is a wrong in the Fig.3 (c) of the ECCV version. The second and third BConv3 in the first column of the figure should be BConv5 and BConv7 respectively.

6. Benchmark RGB-D SOD

The complete RGB-D SOD benchmark can be found in this page:

http://dpfan.net/d3netbenchmark/

7. Acknowledgement

We implement this project based on the code of ‘Cascaded Partial Decoder for Fast and Accurate Salient Object Detection, CVPR2019’ proposed by Wu et al.