juancardoso / road-semantic-segmentation

Semantic Segmentation - It label the pixels of a road in images using a Fully Convolutional Network (FCN).

Semantic Segmentation Fully Convolutional Neural Network

Semantic Segmentation - It label the pixels of a road in images using a Fully Convolutional Network (FCN).

Index

• Prerequisites
• Neural network archtecture
• Some results
• Setup
  • Frameworks and Packages
  • Dataset
  • Add a custom Dataset
• Start
  • Usage
  • google colabotory - GPU 4free

Prerequisites

    pip install opencv-python opencv-contrib-python scipy==1.2.0 Pillow matplotlib tensorflow-gpu numpy colored

Neural network archtecture

The project start loading a vgg model so implements an FCN - Fully Convolutional Network

More details about how VGG16 is used in this archtechture to converte it to a fully convolutional network, you can read 'this paper'.

Some results

Setup

Frameworks and Packages

Make sure you have the following is installed:

• Python 3
• TensorFlow
• NumPy
• SciPy

Dataset

The project will download the Kitti Road dataset dataset from here. It will extract the files in data folder. This will create the folder data_road with all the training a test images.

Add a custom Dataset

Add images in data/data_road/training/

Example:

• gt_image_2/um_lane_000000.png
• image_2/um_000000.png

Obs: The images in qt_image_2 are not a image with 4 channels (png file), are a jpg.

Start

Usage

Run the following command to run the project:

python main.py

Examples to run the project

python main.py -e 25 -i './data/data_road/training/image_2/*.png' -l './data/data_road/training/gt_image_2/*_road_*.png' -n 2
python main.py -m model/model.ckpt --pred_data_from zed --disable_gpu
python main.py -m model/model.ckpt -V teste.png --pred_data_from image
python main.py -m model/model.ckpt -V teste.mp4 --pred_data_from video

it will run the Neural network with default parameter values, to know more, type python main.py -h

Usage: main.py [options]

Options:
    -h, --help            show this help message and exit
    -i GLOB_TRAINIG_IMAGES_PATH, --glob_trainig_images_path=GLOB_TRAINIG_IMAGES_PATH
                        Path where is yours images to train the model. eg:
                        ./data/data_road/training/image_2/*.png'
    -l GLOB_LABELS_TRAINIG_IMAGE_PATH, --glob_labels_trainig_image_path=GLOB_LABELS_TRAINIG_IMAGE_PATH
                        Path where is yours label images to train the model.
                        eg: ./data/data_road/training/gt_image_2/*_road_*.png
    -r LEARN_RATE, --learn_rate=LEARN_RATE
                        The model learn rate | Default=9e-5
    -n NUM_CLASSES, --num_classes=NUM_CLASSES
                        Number of classes in your dataset | Default value = 2
    -e EPOCHS, --epochs=EPOCHS
                        Number of epochs that FCN will train | Default=25
    -b BATCH_SIZE, --batch_size=BATCH_SIZE
                        Number of batch size for each epoch. | Default=4
    -t DATA_PATH, --data_path=DATA_PATH
                        Training data path. | Default='data_road/training'
    -p LOG_PATH, --log_path=LOG_PATH
                        Path to save the tensorflow logs to TensorBoard |
                        Default='.'
    -v VGG_DIR, --vgg_dir=VGG_DIR
                        Path to dowloand vgg pre trained weigths. |
                        Default='./data/vgg'
    -g GRAPH_VISUALIZE, --graph_visualize=GRAPH_VISUALIZE
                        create a graph image of the FCN archtecture. |
                        Default=False
    -m path_data, --path_model
                        Load a model, to predict a data. | 
                        Default=False
    -V PATH_DATA, --path_data", Path to predict a data. | 
                        Default= ''
    --pred_data_from PRED_DATA_FROM, Choose a type predict [video, image, zed] | 
                        Default=video
    --disable_gpu DISABLE_GPU, Disable predict by GPU

Running at google colabotory - GPU 4free

Open this link