TensorFlow implementation of D-LinkNet for road extraction.
Details can be found in this paper:
This model uses ResNet 50 provided by TensorFlow-Slim as encoder. See setting up section for more information.
Dataset is from DeepGlobe Road Extraction Challenge.
- Python 3.5
- CUDA 9.0
- TensorFlow 1.10
Before training this model, download net folder from
https://github.com/tensorflow/models/tree/master/research/slim
and place in the root directory of this project.
This contains necessary files to construct the Res50 model.
Run python ./train_slim_model.py
Options:
--data_dir=<path>
Path to training data.
Satellite images should have names like *sat*, labeled images should have names like *mask*.
--summary_dir=<path>
Save summary to specified path.
Default to `./summary/`
--save_dir=<path>
Save model to specified path.
By default the model will be saved under `<path>/model_<time_string>/`.
Default to `./model`.
--no_append
If set, model will be saved directly under `save_dir`. No sub directory will be made.
--resume_dir=<path>
If set, resume training the model from a previous checkpoint.
--CKPT_RES50=<path>
Path to ResNet 50 pre-trained model.
Default to `./pretrained-checkpoint/resnet_v1_50.ckpt1`.
--num_epoch=<int>
Specify number of epochs to train. Default to 16.
--partial_train
If set, parameters in Res50 model will not be updated.
Run python ./test_slim_model.py
Options:
--input_dir=<path>
Path to test files.
Satellite image files should have name like `_sat.*`.
Label images (if present) should have name like `_mask.*`.
If label images are present, iou and loss will be computed.
--output_dir=<path>
Path to save results.
If set, save prediction files to <path>.
Otherwise results will be saved to input_dir
--ckpt_dir=<path>
Path to saved model (checkpoint files).
--pb_dir=<path>
If set, load model from frozen graph. This option overrides `ckpt_dir`.
Run python ./freezer.py
Options:
--ckpt_dir=<path>
Path to checkpoint files.