README.md

Synthetic Segmentation Dataset Generator Using Realistic City Backgrounds

A project to generate synthetic dataset with realistic city backgrounds for image segmentation problems.

If you use this code in your research project, please cite our SİU 2022 paper:

@INPROCEEDINGS{sahibindensiu2022,
  author={İşlek, İrem and Aksaylı, N. Deniz and Güngör, Onur and Karaman, Çağla Çığ},
  booktitle={2022 30th Signal Processing and Communications Applications Conference (SIU)},
  title={Generating Synthetic Image Segmentation Dataset Using Realistic City Backgrounds},
  year={2022}
  }

Setup - Install requirements

pipenv install -r requirements.txt

For Background Images Dataset Generation

Note: This step is needed if you want to generate background images from a mp4 video.
Please define the source video path for background images and the folder name
for output background images using the dataset_generator_parameters.ini file.

pipenv run python background_images_generator.py

For Object/Mask Dataset

Note: This step is needed if you want to use the pedestrian objects from the PennFudan Dataset.
Otherwise assign the related path of your objects/masks dataset to the "segmentation_object_image_folder"
and "segmentation_object_mask_folder" variables in the dataset_generator_parameters.ini file.

pipenv run python PennFudan_mask_generator.py

For Synthetic Dataset Generation (Merging Background Images and Objects/Masks)

Note: This step merges background images with the object images and related image masks to generate a synthetic segmentation dataset.
Please check synthetic_segmentation_dataset_generator_params from the dataset_generator_parameters.ini file
to customize your synthetic dataset.

pipenv run python synthetic_segmentation_dataset_generator.py

Parameters for Customization

These parameters can be changed from dataset_generator_parameters.ini file

frame_delay: determines the delay of taking frames from the background video (in seconds)
mask_image_reuse_count: determines how many synthetic samples are created for each mask
object_min_height_threshold: determines a threshold for using objects which has min n pixel height
max_object_count_per_image: determines the maximum object count per image