Fully Convolutional Networks for Portrait Matting

Basic Knowledges

• Fully Convolutional Networks (Chinese)

Referenced Papers

• Fully Convolutional Networks for Semantic Segmentation
• Automatic Portrait Segmentation for Image Stylization

Referenced Repositories

• shekkizh/FCN.tensorflow
• PetroWu/AutoPortraitMatting

Data Sets

• MIT Scene Parsing ADE 20K
  • 20210 training images (not used)
  • 2000 validation images
• Portraits 2K
  • 1719 training images (used)
• PASCAL VOC 2012 3K
  • 2913 training images (used)

Model

• conv1_1~conv5_3: VGG-19
• pool5~conv_t3: Convolution (replace dense network) and Deconvolution

Directory Structure

.Fully-Convolutional-Networks/
├── .Data_zoo/
│   └── .MIT_SceneParsing/
│       ├── ADEChallengeData2016.zip
│       ├── MITSceneParsing.pickle
│       ├── train_data.npz (show up after first train)
│       ├── val_data.npz (show up after first train)
│       └── .ADEChallengeData2016/
│           ├── sceneCategories.txt
│           ├── .images/
│           │   ├── .training/
│           │   └── .validation/
│           │
│           └── .annotations/
│               ├── .training/
│               └── .validation/
│
├── .Model_zoo/
│   └── imagenet-vgg-verydeep-19.mat
│
├── .logs/
│   ├── checkpoint
│   ├── model.ckpt-100000.data-00000-of-00001
│   ├── model.ckpt-100000.meta
│   └── model.ckpt-100000.index
│
├── fcn.py (main program)
├── augment.py
├── batch_datset_reader.py
├── reader.py
├── tensorflow_utils.py
└── README.md

• For each data, the filename in image/, annotation/ folder must be same.

Data Augmentations

• Flip: 50% horizontally
• Rotation: -90 ~ +90
• Scale: 0.5 ~ 1.5
• Shift: -50% ~ +50% horizontally & vertically

Requirements

• tensorflow-gpu == 1.2.1

Quick Usage

• Train
  • python3.5 fcn.py -m train
• Visualize
  • python3.5 fcn.py -m visualize
• Test
  • python3.5 fcn.py -m test -tl <test_list>
• To see full usage
  • python3.5 fcn.py --help