We submit it to Experiments in Fluids. Some useful information is also provided here.

1. Directory structure
2. Installation guidance
3. More experimental results
4. Some tips about our method
5. Links to data generaton,experiment arrangement

├── PIV-DCNN20170710
│   ├── readme.md
│   ├── install.m
│   ├── data
│   │   ├── ImagesForDataset
│   │   │   ├── readme.md
│   │   │   └── A1-10.bmp
│   │   ├── NetF1~F4-3
│   │   │   ├── imdb.mat
│   │   │   ├── net-epoch-1.mat
│   │   │   └── net-train.pdf
│   │   └── TestImages
│   │       ├── readme.md
│   │       └── *.tif, *bmp, *jpg
│   ├── PIV_DNN
│   │   ├── readme.md
│   │   ├── ChangesForMatConvNet
│   │   │   └── some files 
│   │   ├── cnn_pivdnn_init.m 
│   │   ├── cnn_pivdnn.m
│   │   ├── genPIVImgDB_F4.m
│   │   ├── genPIVImgDB.m
│   │   ├── ImgData.mat
│   │   ├── Nets.mat
│   │   ├── pivdnn.m
│   │   └── train.m
│   ├── experiments
│   │   ├── readme.md
│   │   └── exp0-7.m
│   ├── matconvnet
│   │   └── MatConvNet files
│   ├── OpticalFlow
│   │   └── OpticalFlow files
│   ├── PIVCC
│   │   ├── dctn.m
│   │   ├── idctn.m
│   │   ├── PIV_analysis.m
│   │   └── smoothn.m
│   └── pivlab
│   │   └──PIVlab files

• Install GPU driver, CUDA and CuDNN.
• Run install.m in the main folder. It will download the MatConvNet, compile files, and copy some customized files to the MatConvNet automatically.
• Cautions: You need to specify the configure terms (enable GPU, CUDA root, ...) in install.m if the default values do not work. (Note, the slow CPU version without GPU support is also ok)
• Enjoy it. cd '../PIV_DNN'; run('PIVdnn.m'); or cd '../experiments';run('exp1.m').

Information about the packages

• PIVlab website
• MatConvNet website

• Raw Images

  

• Vector field (color contour reprensents corresponding magnitude)

  

• Vector component histogram

  

• Raw Images

  

• Vector field (color contour reprensents corresponding magnitude)

  

• Smoothed vector field

  

• Vector component histogram

  

• Raw Images

  

• Vector field (color contour reprensents corresponding magnitude)

  

• Vector component histogram

  

• System requirement(8G RAM, Hard disk space > 50G, GPU with computation ability >3.0 is promotional)
• CuDNN is highly recommended
• Training 6 DCNN nets(optional, about 2 days with GPU accelaration):cd '../PIV_DNN';run('train.m')
• The figures in our manuscript (about 2 day to run massive Monte - Carlo simulation):cd '../experiments'; run('exp0.m')

Yong Lee (Email: [email protected]) @2017-07-10