tomo0318/segmentation_unet

Semantic segmentation using U-NET

Semantic Segmentation using U-Net on Pascal VOC 2012

This repository implements semantic segmentation on Pascal VOC2012 using U-Net.

An article about this implementation is here.

Semantic segmentation is a kind of image processing as below.

This package includes modules of data loader, reporter(creates reports of experiments), data augmenter, u-net model, and training it.

Usage

To show how to run.

python main.py --help

To run with data augmentation using GPUs.

python main.py --gpu --augmentation

U-Net

U-Net is an encoder-decoder model consisted of only convolutions, without fully connected layers.

U-Net has a shape like "U" as below, that's why it is called U-Net.

Experiments

The following results is got by default settings.

Results of segmentation

For the training set

For the test set

Accuracy and Loss