keras-image-segmentation
keras-image-segmentation copied to clipboard
Image segmentation with keras. FCN, Unet, DeepLab V3 plus, Mask RCNN ... etc.
Keras Image Segmentation
Semantic Segmentation easy code for keras users.
We use cityscape dataset for training various models.
Use pretrained VGG16 weight for FCN and U-net! You can download weights offered by keras.
Tested Env
- python 2 & 3
- tensorflow 1.5
- keras 2.1.4
- opencv 3.3
File Description
File | Description |
---|---|
train.py | Train various models. |
test.py | Predict one picture what you want. |
dataest_parser/make_h5.py | Parse cityscape dataset and make h5py file. |
dataest_parser/generator.py | Data_generator with augmentation using data.h5 |
model/ | Folder that contains various models for semantic segmentation |
segmentation_dh/ | Experiment folder for Anthony Kim(useless contents for users) |
segmentation_tk/ | Experiment folder for TaeKang Woo(useless contents for users) |
temp/ | Folder that contains various scripts we used(useless contents for users) |
Implement Details
We used only three classes in the cityscape dataset for a simple implementation.
Person, Car, and Road.
Simple Tutorial
First, you have to make .h5 file with data!
python3 dataset_parser/make_h5.py --path "/downloaded/leftImg8bit/path/" --gtpath "/downloaded/gtFine/path/"
After you run above command, 'data.h5' file will appear in dataset_parser folder.
Second, Train your model!
python3 train.py --model fcn
Option | Description |
---|---|
--model | Model to train. ['fcn', 'unet', 'pspnet'] |
--train_batch | Batch size for train. |
--val_batch | Batch size for validation. |
--lr_init | Initial learning rate. |
--lr_decay | How much to decay the learning rate. |
--vgg | Pretrained vgg16 weight path. |
Finally, test your model!
python3 test.py --model fcn
Option | Description |
---|---|
--model | Model to test. ['fcn', 'unet', 'pspnet'] |
--img_path | The image path you want to test |
Todo
- [x] FCN
- [x] Unet
- [x] PSPnet
- [ ] DeepLab_v3
- [ ] Mask_RCNN
- [ ] Evauate methods(calc mIoU)
Contact us!
Anthony Kim: [email protected]
TaeKang Woo: [email protected]