Keras implementation of semantic segmentation FCNs
Update Logs
2019-02-22: implemented several common FCNs and support Geo-tiff Images (especially for remote sensing images).
2019-03-07: tested on VOC2012 dataset (multi-class, rgb inputs) and Inria dataset (binary class, rgb inputs).
TODO
- More SOTA FCN architectures.
- Support different output strides in backbones.
- Support different open data sets like VOC, CityScapes, ADE20K, MSCOCO, etc.
- More flexible in data format.
Backbones(Encoders) that have been implemented
- ResNet_v2 (including ResNet_v2-50, ResNet_v2-101, ResNet_v2-150, ResNet_v2-200)
- ResNet_v2_separable (including ResNet_v2-50, ResNet_v2-101, ResNet_v2-150, ResNet_v2-200)
- VGG (inclufing VGG-16, VGG-19)
- Xception-41
Backbones to be implemented
- DenseNet
FCNs that have been implemented
- FCN (including FCN-8s, FCN-16s, FCN-32s)
- SegNet
- U-Net, Res U-Net, Mobile U-Net
- PSPNet
- RefineNet
- Deeplab v3
- Deeplab v3+
- Dense ASPP
FCNs to be implemented
- ICNet
- etc
.
├── core
│ ├── __init__.py
| |
│ ├── configures.py
| |
│ ├── encoder
│ | ├── __init__.py
│ | ├── resnet_v2.py
│ | ├── resnet_v2_separable.py
│ | ├── vggs.py
│ | └── xceptions.py
│ |
│ ├── nets
│ | ├── __init__.py
│ | ├── deeplabs.py
│ | ├── dense_aspp.py
│ | ├── fcns.py
│ | ├── pspnets.py
│ | ├── refinenets.py
│ | ├── segnets.py
│ | └── unets.py
│ |
│ └── utils
│ ├── __init__.py
│ |
│ ├── data_utils
│ | ├── __init__.py
│ | ├── image_io_utils.py
│ | ├── label_transform_utils.py
│ | ├── image_augmentation_utils.py
| | ├── directory_iterator.py
│ | └── data_generator.py
│ |
│ ├── loss_utils.py
│ ├── metric_utils.py
│ ├── model_utils.py
│ ├── net_utils.py
│ ├── training_utils.py
│ ├── predicting_utils.py
│ └── vis_utils.py
│
├── data
│
├── examples
│ ├── __init__.py
│ ├── s1_generate_datasets.py
│ ├── s2_training.py
| ├── directory_iterator.py
│ └── s3_predicting.py
│
├── tools
| ├── color2index.py
| ├── generate_dataset.py
│ └── helpers.py
|
├── LICENSE
|
└── README.md
The source code was compiled in a Windows 10 platform using Python 3.6. The dependencies include:
tensorflow-gpu: 1.9, backend
Keras: 2.2.4, framework
opencv: 4.0, used for image I/O
PIL: used for image I/O
numpy: used for array operations
matplotlib: used to visualize images
tqdm: used to log iterations
GDAL: used for geo-spatial image I/O
scikit-learn: used for metric evaluation
- Since we can randomly crop patch images from the input images, we do not need to prepare training images.
- However, for large images like HSR remote sensing image tiles, cropping patch images when training is extraordinarily time-consuming.
- Thus for such datasets, we can crop them into a certain size and resize the input images to target size for convenience.
- More can be found in
./tools/generate_dataset.py
- More can be found in
./tools/color2index.py
The following figure shows converting colors to labels:
The following figure shows converting labels to colors:
dataset are organized as the following folder structure:
voc
├── image # input image
├── label_color # colorful label
├── label_index # gray indexed label
├── prediction # prediction
├── _temp # temp dir
├── models # model instance
├── logs # training logs
├── train.txt # training file names
└── val.txt # validation file names
- WHU Building Dataset [Data] [Paper]
- Inria Aerial Building Labeling Dataset [Data] [Paper]
- ISPRS 2D Semantic Labeling Benchmark [Main]
- Massachusetts Roads and Buildings Dataset [Data]
- VOC2012 [Main] [Data]
- VOC2012 Augmentation [Data]
- CityScapes Dataset [Data]
- ADE20K Dataset [Data]
Penghua Liu (liuph3@mail2.sysu.edu.cn), Sun Yat-sen University