JL-DCF-pytorch

Pytorch implementation for JL-DCF: Joint Learning and Densely-Cooperative Fusion Framework for RGB-D Salient Object Detection (CVPR2020) [PDF][中文版], Siamese Network for RGB-D Salient Object Detection and Beyond (TPAMI2021) [PDF][中文版]

Requirements

Python 3.6
Pytorch 1.5.0
Torchvision 0.6.1
Cuda 10.0

Usage

This is the Pytorch implementation of JL-DCF. It has been trained and tested on Windows (Win10 + Cuda 10 + Python 3.6 + Pytorch 1.5), and it should also work on Linux but we didn't try.

To Train

Download the pre-trained ImageNet backbone (resnet101/resnet50, densenet161, vgg16 and vgg_conv1, whereas the latter already exists in the folder), and put it in the 'pretrained' folder.
Download the training dataset and modify the 'train_root' and 'train_list' in the main.py.
Start to train with

python main.py --mode=train --arch=resnet --network=resnet101 --train_root=xx/dataset/RGBDcollection --train_list=xx/dataset/RGBDcollection/train.lst

The converged loss value is around 7000.

To Test

Download the testing dataset and have it in the 'dataset/test/' folder.
Download the already-trained JL-DCF pytorch model and modify the 'model' to its saving path in the main.py.
Start to test with

python main.py --mode=test --arch=resnet --network=resnet101 --model=xx/JLDCF_resnet101.pth --sal_mode=LFSD  --test_folder=test/LFSD

Be careful that model and network should match.

Learning curve

The training log is saved in the 'log' folder. If you want to see the learning curve, you can get it by using: tensorboard --logdir your-log-path

Pre-trained ImageNet model for training

densenet161
resnet101
resnet50
vgg16
vgg_conv1, password: rllb

Trained model for testing

Resnet101:

Baidu Pan: resnet101, password: jdpb
Google Drive: https://drive.google.com/open?id=12u37yz-031unDPJoKaZ0goK8BtPP-6Cj

Resnet50:

Baidu Pan: resnet50, password: ve9z

vgg16:

Baidu Pan: vgg16, password: 4yv6

Densenet161:

Baidu Pan: densenet161, password: qza4

JL-DCF-pytorch saliency maps

Resnet101 (NJU2K, NLPR, STERE, RGBD135, LFSD, SIP, DUT-RGBD):

Baidu Pan: resnet101, password: v144
Google Drive: https://drive.google.com/file/d/1GoqDlLrN_INNldsbOEAxyyu1QB_urHr0/view?usp=sharing

Resnet50 (NJU2K, NLPR, STERE, RGBD135, LFSD, SIP, DUT-RGBD):

Baidu Pan: resnet50, password: zen9

vgg16 (NJU2K, NLPR, STERE, RGBD135, LFSD, SIP, DUT-RGBD):

Baidu Pan: vgg16, password: cuui

Densenet161 (NJU2K, NLPR, STERE, RGBD135, LFSD, SIP, DUT-RGBD):

Baidu Pan: densenet161, password: ovuc

Important Tips!

Note that our JL-DCF model was trained on depth maps which satisfy the rule that closer objects present lower depth values (are "black"), while further objects have higher depth values (are "white"). Although these depth maps are normalized into the range [0, 1], such a rule is enforced to meet physical common sense. We observed that the model performance would somewhat degrade when using reversed maps during testing. So be aware of this issue when testing the models. Anyway, the training and testing settings should always be kept consistent.

Dataset

Baidu Pan:
Training dataset (with horizontal flip), password: i4mi
Testing datadet, password: 1ju8
Google Drive:
Training dataset (with horizontal flip)
Testing datadet

Performance

Below is the performance of JL-DCF-pyotrch (Pytorch implementation). Generally, the performance of Pytorch implementation is comparable to, and even slightly better than the previous Caffe implementation reported in the paper. This is probably due to the differences between deep learning platforms. Also, due to the randomness in the training process, the obtained results will fluctuate slightly.

Datasets	Metrics	Pytorch ResNet101	Pytorch ResNet50	Pytorch VGG16	Pytorch DenseNet161
NJU2K	S-measure	0.917	0.913	0.910	0.917
	maxF	0.919	0.915	0.912	0.917
	maxE	0.950	0.951	0.949	0.952
	MAE	0.037	0.039	0.038	0.037
NLPR	S-measure	0.931	0.931	0.926	0.934
	maxF	0.920	0.918	0.915	0.924
	maxE	0.964	0.965	0.963	0.967
	MAE	0.022	0.022	0.024	0.020
STERE	S-measure	0.906	0.900	0.900	0.909
	maxF	0.903	0.895	0.898	0.905
	maxE	0.946	0.942	0.942	0.947
	MAE	0.040	0.044	0.042	0.039
RGBD135	S-measure	0.934	0.928	0.925	0.934
	maxF	0.928	0.918	0.918	0.926
	maxE	0.967	0.957	0.960	0.964
	MAE	0.020	0.021	0.021	0.020
LFSD	S-measure	0.862	0.850	0.833	0.863
	maxF	0.861	0.855	0.839	0.868
	maxE	0.894	0.887	0.879	0.900
	MAE	0.074	0.081	0.084	0.073
SIP	S-measure	0.879	0.885	0.887	0.894
	maxF	0.889	0.894	0.896	0.903
	maxE	0.925	0.931	0.931	0.934
	MAE	0.050	0.049	0.047	0.044
DUT-RGBD	S-measure	0.905	0.894	0.881	0.914
	maxF	0.903	0.892	0.878	0.916
	maxE	0.937	0.928	0.921	0.945
	MAE	0.042	0.048	0.054	0.040

Citation

Please cite our paper if you find the work useful:

@inproceedings{Fu2020JLDCF,
title={JL-DCF: Joint Learning and Densely-Cooperative Fusion Framework for RGB-D Salient Object Detection},
author={Fu, Keren and Fan, Deng-Ping and Ji, Ge-Peng and Zhao, Qijun},
booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
pages={3052--3062},
year={2020}
}
    
@article{Fu2021siamese,
title={Siamese Network for RGB-D Salient Object Detection and Beyond},
author={Fu, Keren and Fan, Deng-Ping and Ji, Ge-Peng and Zhao, Qijun and Shen, Jianbing and Zhu, Ce},
journal={IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)},
year={2021}
}

Benchmark RGB-D SOD

The complete RGB-D SOD benchmark can be found in this page
http://dpfan.net/d3netbenchmark/

Estia-xmy / JL-DCF-pytorch