- Fully Convolutional Networks (Chinese)
- Fully Convolutional Networks for Semantic Segmentation
- Automatic Portrait Segmentation for Image Stylization
- 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)
conv1_1~conv5_3: VGG-19pool5~conv_t3: Convolution (replace dense network) and Deconvolution
.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.
- Flip: 50% horizontally
- Rotation: -90 ~ +90
- Scale: 0.5 ~ 1.5
- Shift: -50% ~ +50% horizontally & vertically
tensorflow-gpu == 1.2.1
- Train
python3.5 FCN.py -m train
- Visualize
python3.5 FCN.py -m visualize
- Test
python3.5 FCN.py -m test -tl <test_list>
- Mat video
./mat.sh <video_name>
- To see full usage
python3.5 FCN.py --help
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent