footballfields.ini

Contents:

# This config file contains the specific settings for this orthoseg project.
#
# The config used for an orthoseg project is loaded in the following order:
#   1) the project defaults as "hardcoded" in orthoseg (project_defaults.ini)
#   2) any .ini files specified in the general.extra_config_files_to_load 
#      parameter (in this file).
#   3) this config file
# Parameters specified in a config file loaded later in the order above
# overrule the corresponding parameter values specified in a previously 
# loaded config file.

# General settings.
[general]

# Extra config files to load for this project. They will be loaded in the 
# order specified and can be specified one path per line, comma seperated.
# If a relative path is used it will be resolved towards the parent dir of 
# this file.
extra_config_files_to_load = ../project_defaults_overrule.ini

# The subject that will be segmented.
segment_subject = footballfields

# Settings concerning the neural network model you want to use for the 
# segmentation. 
[model]

# The segmentation architecture to use. 
# 
# The architectures currently supported by orthoseg follow the encoder-decoder 
# principle:
#   * an encoder: a (deep) neural network that detects features on object level
#   * a decoder: a (deep) neural network that converts the detected features 
#     on object level to a segmentation on pixel level
#
# To configure an encoder/decoder architecture in orthoseg, specify the it in 
# the following way: architecture = {encoder}+{decoder}, eg. the default is:
#   architecture = inceptionresnetv2+unet
#
# As this is just a sample project, use a light-weight architecture.
architecture = mobilenetv2+linknet

# Settings concerning the train process.
[train]

# Parameters regarding the size/resolution of the images used to train on.
#
# The size the label_location boxes need to be digitized depends on these values:
# e.g. with image_pixel_width = 512 and image_pixel_x_size = 0.25, the boxes need to be
# 512 pixels * 0.25 meter/pixel = 128 meter wide.
#
# For some model architectures there are limitations on the image sizes
# supported. E.g. if you use the linknet decoder, the images pixel width and height
# has to be divisible by factor 32.
image_pixel_width = 512
image_pixel_height = 512
image_pixel_x_size = 0.25
image_pixel_y_size = 0.25

# In json format, the classes to train/predict and for each class:
#     * the label names in the training data to use for this class  
#     * the weight to use when training
classes =   {   "background": {
                    "labelnames": ["ignore_for_training", "background"],
                    "weight": 1
                },
                "footballfield": {
                    "labelnames": ["footballfield"],
                    "weight": 1
                }
            }

# Settings concerning the prediction process.
[predict]

# The batch size to use. 
# Depends on available hardware, model used and image size.
batch_size = 1

# Parameters regarding the size/resolution of the images to run predict on.
#
# For some model architectures there are limitations on the image sizes
# supported. E.g. if you use the linknet decoder, the images pixel width and height
# has to be divisible by factor 32.
image_pixel_width = 1024
image_pixel_height = 1024
image_pixel_x_size = 0.25
image_pixel_y_size = 0.25
image_pixels_overlap = 128