TensorFlow-based semantic segmentation codes.

Implemented models and data I/Os

Model

• Fully Convolutional Network model.
• Dilated Network model.
• Pyramid Scene Parsing Network model.

Data I/O

• Cityscapes dataset.
• VOC2012 dataset.

Prepare .tfrecord data

This repo only assumes .tfrecord format data.
The .tfrecord data should be created by scripts/write_tfrecord_${data_name}.py.

Training

You can train model by scripts/trainer_${model_name}_${data_name}.py.
Example parameters are defined in scripts/params/train_params_${model_name}_${data_name}.yaml.

Prediction

You can run model inference by scripts/predictor_${model_name}_${data_name}.py.
Example parameters are defined in scripts/params/predict_params_${model_name}_${data_name}.yaml.