frequencyxxq/dmra

Code and Dataset for ICCV 2019 paper. "Depth-induced Multi-scale Recurrent Attention Network for Saliency Detection".

DMRA_RGBD-SOD

Code repository for our paper entilted "Depth-induced Multi-scale Recurrent Attention Network for Saliency Detection" accepted at ICCV 2019 (poster).

Overall

The proposed Dataset

• Dataset: DUTLF

1. This dataset consists of DUTLF-MV, DUTLF-FS, DUTLF-Depth.
2. The dataset will be expanded to 3000 about real scenes.
3. We are working on it and will make it publicly available soon.

• Dataset: DUTLF-Depth

1. The dataset is part of DUTLF dataset captured by Lytro camera, and we selected a more accurate 1200 depth map pairs for more accurate RGB-D saliency detection.
2. We create a large scale RGB-D dataset(DUTLF-Depth) with 1200 paired images containing more complex scenarios, such as multiple or transparent objects, similar foreground and background, complex background, low-intensity environment. This challenging dataset can contribute to comprehensively evaluating saliency models.

• The dataset link can be found here. And we split the dataset including 800 training set and 400 test set.

DMRA Code

> Requirment

• pytorch 0.3.0+
• torchvision
• PIL
• numpy

> Usage

1. Clone the repo

git clone https://github.com/jiwei0921/DMRA.git
cd DMRA/

2. Train/Test

• test
  Download related dataset link, and set the param '--phase' as "test" and '--param' as 'True' in demo.py. Meanwhile, you need to set dataset path and checkpoint name correctly.

python demo.py

• train
  Our train-augment dataset link [ fetch code haxl ] / train-ori dataset, and set the param '--phase' as "train" and '--param' as 'True'(loading checkpoint) or 'False'(no loading checkpoint) in demo.py. Meanwhile, you need to set dataset path and checkpoint name correctly.

python demo.py

> Training info and pre-trained models for DMRA

To better understand, we retrain our network and record some detailed training details as well as corresponding pre-trained models.

Iterations	Loss	NJUD(F-measure)	NJUD(MAE)	NLPR(F-measure)	NLPR(MAE)	download link
100W	958	0.882	0.048	0.867	0.031	link
70W	2413	0.876	0.050	0.854	0.033	link
40W	3194	0.861	0.056	0.823	0.037	link
16W	8260	0.805	0.081	0.725	0.056	link
2W	33494	0.009	0.470	0.030	0.452	link
0W	45394	-	-	-	-	-

• Tips: The results of the paper shall prevail. Because of the randomness of the training process, the results fluctuated slightly.

> Results

| DUTLF-Depth | | NJUD | | NLPR | | STEREO | | LFSD | | RGBD135 | | SSD |

• Note: For evaluation, all results are implemented on this ready-to-use toolbox.

> Related RGB-D Saliency Datasets

All common RGB-D Saliency Datasets we collected are shared in ready-to-use manner.

• The web link is here.

If you think this work is helpful, please cite

@InProceedings{Piao_2019_ICCV,       
   author = {Yongri {Piao} and Wei {Ji} and Jingjing {Li} and Miao {Zhang} and Huchuan {Lu}},   
   title = {Depth-induced Multi-scale Recurrent Attention Network for Saliency Detection},     
   booktitle = "ICCV",     
   year = {2019}     
}  

Contact Us

If you have any questions, please contact us ( [email protected] ).