Tensorflow - Semantic Segmentation Neural Network

An implementation of Convolutional Neural Network for Semantic Segmentation task.

This is a deep network of my own design named KNetS, inspired by the [DeepLab v.2 Network] (https://arxiv.org/abs/1606.00915)

Requirements

In addition to Python 3, Tensorflow Version>= 1.3, and Matlab the following packages are required:

numpy scipy pillow matplotlib

These packages can be installed by running pip3 install -r requirements.txt or pip3 install numpy scipy pillow matplotlib.

Usage

Currently, the code to generate the data was written to work only with the Camvid dataset.

Step 1: Creating .mat file from raw data

Download [Camvid dataset] (http://mi.eng.cam.ac.uk/research/projects/VideoRec/CamVid/)

Put images data to the folder ./matlab/Camvid/701_StillsRaw_full.

Put labels data to the folder ./matlab/Camvid/701_Label.

Use Maltab to run make_mat_files.m in the matlab folder to create .mat data files.

Modify the img_size and partition rate if needed.

Step 2: Creating tfrecord file from .mat file

Run python3 data_utils.py to create tfrecord files.

If needed, the user can use the showing_data_from_tfrecord() function in the data_utils.py to visualize the generated data in tfrecord files.

Step 3: Training the network

Run python3 main to train the network.

Modify batch_size, learning_rate and num_record if needed.

After running the saved_model folder will be created.

Monitor the learning process by using terminal and navigate to the graph folder inside the saved_model folder, then run tensorboard --logdir=KNetS to activate Tensorboard.

The training process can be resume if interrupted if the saved model has been saved. Just run python3 main again.

Step 4: Testing the network

Put testing images in a folder and use the function test_segmentize() in the data_utils.py to segmentize the test samples

Model

• KNetS
• FCN8