Code repository for our paper entilted "Depth-induced Multi-scale Recurrent Attention Network for Saliency Detection" accepted at ICCV 2019 (poster).
- Dataset: DUTLF
- This dataset consists of DUTLF-MV, DUTLF-FS, DUTLF-Depth.
- The dataset will be expanded to 3000 about real scenes.
- We are working on it and will make it publicly available soon.
- Dataset: DUTLF-Depth
- 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.
- 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.
- pytorch 0.3.0+
- torchvision
- PIL
- numpy
git clone https://github.com/jiwei0921/DMRA.git
cd DMRA/
- test
Download related dataset link, and set the param '--phase' as "test" and '--param' as 'True' indemo.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) indemo.py
. Meanwhile, you need to set dataset path and checkpoint name correctly.
python demo.py
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.
| DUTLF-Depth | | NJUD | | NLPR | | STEREO | | LFSD | | RGBD135 | | SSD |
- Note: For evaluation, all results are implemented on this ready-to-use toolbox.
All common RGB-D Saliency Datasets we collected are shared in ready-to-use manner.
- The web link is here.
@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}
}
If you have any questions, please contact us ( [email protected] ).