Awesome

MAC-light-field-saliency-net

We release the code of Light Field Saliency Detection with Deep Convolutional Networks, built on the DeepLab v2-Caffe codebase.

Overall network

Architectures of three proposed MAC block variants

Installation

Python 2.7

Caffe

Matlab 2016b

Do the following two steps to make sure that .m files can be used in Python:

• cd matlabroot/extern/engines/python
• python setup.py

Note: matlabroot is the root directory of MATLAB in your system.

Data

Dataset：Lytro Illum

This project introduces a new light field dataset using a Lytro Illum camera. We collect 640 light fields with significant variations in terms of size, textureness, background clutter and illumination, etc. We generate micro-lens image arrays and central viewing images, and produce corresponding ground-truth maps. Please download them here or BaiduPan.

Other datasets：

We also provide links to two other datasets for light field saliency detection.

|LFSD | |HFUT-Lytro|

Data pre-processing

• Put the micro-lens image arrays into data/original_data/ and ground-truths into data/original_GT/.

• Run python augment.py for data augmentation.

• Convert GTs to the labels that can be entered into the Caffe network by running

  python convert_labels.py ./data/train/GT_aug/ ./data/train/name.txt ./data/train/annotations/

• Copy the labels to the val/annotations/: cp ./data/train/annotations/* ./data/val/annotations/

• Copy the original images to the val/JPGImages/: cp ./data/original_data/* ./data/val/JPGImages/

• Put training and testing data in data/train/ and data/val/. 5-fold cross-validation is used in the project. The generated image index are in the .txt files. For example, train1.txt indicates the train data used in the 1st-fold experiment.

Train

• Replace the caffe_root in deeplab-public-ver2/colorname_layers.py with the caffe path in your system.

• We initialize the backbone model with DeepLabv2 pre-trained on the PASCAL VOC 2012 segmentation benchmark. Then Create a new "pretrain" folder and place the downloaded pretrained model in this folder.

• There are two ways to train the whole network: (1) Run python train.py to train the MAC block-9×9 network. If you need to train other models, you can replace MACblock_99 with MACblock_33 or MACblock_StarShaped. k indicates the first fold cross validation, which can be changed into 1-5. (2) Open the terminal, switch to the caffe root directory, run: /models/MACblock_99/train_LF_net.sh. Note that the file path is changed to the path of your system.

Pretrained models

To get the pretrained models on the Lytro Illum dataset, please download them form GoogleDrive or BaiduPan. Then, please put them under the corresponding models paths.

Test

• Run python test.py to test and save the test results as a .mat file.

• Run /evaluation/saliencymap.m to get the final saliency map. The other .m files in evaluation are used to calculate quantitative indicators such as F-measure, weighted F-measure, mean absolute error, average precision, and precision-recall curve.

Results

The computed saliency maps of three proposed models on the Lytro illum dataset are provided in the repo.

Visual comparison of our best MAC block variant (MAC block-9 × 9) and state-of-the-art methods on three datasets is shown in the following figure. (a) Central viewing/all-focus images. (b) Ground truth maps. (c) MAC block-9 × 9. (d) LFS [1]. (e) DILF [2]. (f) WSC [3]. (g) Multi-cue [4]. The first five samples are taken from the proposed Lytro Illum dataset, the middle three samples are taken from the HFUT-Lytro dataset, and the last two samples are taken from the LFSD dataset.

References

1. N. Li, J. Ye, Y. Ji, H. Ling, and J. Yu, “Saliency detection on light field,” in IEEE Conference on Computer Vision and Pattern Recognition, 2014.
2. J. Zhang, M. Wang, J. Gao, Y. Wang, X. Zhang, and X. Wu, “Saliency detection with a deeper investigation of light field,” in International Joint Conference on Artificial Intelligence, 2015.
3. N. Li, B. Sun, and J. Yu, “A weighted sparse coding framework for saliency detection,” in IEEE Conference on Computer Vision and Pattern Recognition, 2015.
4. J. Zhang, M. Wang, L. Lin, X. Yang, J. Gao, and Y. Rui, “Saliency detection on light field: A multi-cue approach,” ACM Transactions on Multimedia Computing, Communications, and Applications, 2017.

TODO

• Raw data release

Citation

If you find our paper and repo useful, please cite our paper. Thanks!

@article{Zhang2020,
    title={Light Field Saliency Detection with Deep Convolutional Networks},
    author={Zhang, Jun and Liu, Yamei and Zhang, Shengping, Poppe, Ronald and Wang, Meng},
    journal={IEEE Transactions on Image Processing},
    volume={29},
    pages={4421-4434}，
    year={2020}
}  

Acknowledgement

Our light field saliency network is built on the DeepLab v2-Caffe, but it could be extended to other network architectures. We thank Liang-Chieh Chen for releasing DeepLabv2-Caffe codebase.

Contact

Jun Zhang, Yamei Liu

Questions can also be left as issues in the repository. We will be happy to answer them.