Synthetic Data Generation

A bunch of scripts to generate synthetic images for YOLO.

Install

1. Install the required packages:

pip install synthetic-data-gen

Tools

Extract Labelme Objects

With this scripts, you can extract objects, which are annotated with labelme, from images.

synthetic-data-gen extract --input_dir INPUT_DIR --output_dir OUTPUT_DIR --margin MARGIN

• INPUT_DIR: The directory where the images are stored.
• OUTPUT_DIR: The directory where the extracted objects will be stored.
• MARGIN: The margin around the object. Usefull to add some space around the object and blend it into the background.

Generate Synthetic Images

With this script, you can generate synthetic images with the extracted objects and corresponding backgrounds.

Minimal example:

synthetic-data-gen generate --input_dir INPUT_DIR --output_dir OUTPUT_DIR --image_number IMAGE_NUMBER

• INPUT_DIR: The directory where the extracted objects are stored.
• OUTPUT_DIR: The directory where the synthetic images will be stored.
• IMAGE_NUMBER: The number of synthetic images to generate.

INPUT_DIR

The input directory should have the following structure:

INPUT_DIR
├── backgrounds
│   ├── background_1.jpg
│   ├── background_2.jpg
│   └── ...
├── foregrounds
|   ├── object_1
|   │   ├── object_1_1.png
|   │   ├── object_1_2.png
|   │   └── ...
|   ├── object_2
|   │   ├── object_2_1.png
|   │   ├── object_2_2.png
|   │   └── ...
|   └── ...
└── labels (optional - use with '--yolo_input')
    ├── background_1.txt
    ├── background_2.txt
    └── ...

Maximal example:

synthetic-data-gen generate --input_dir INPUT_DIR --output_dir OUTPUT_DIR --image_number IMAGE_NUMBER --augmentation_path AUGMENTATION_PATH --max_objects_per_image MAX_OBJECTS_PER_IMAGE --image_width IMAGE_WIDTH --image_height IMAGE_HEIGHT --fixed_image_sizes --scale_foreground_by_background_size --scaling_factors SCALING_FACTORS SCALING_FACTORS --avoid_collisions --parallelize --yolo_input --yolo --color_harmon_alpha COLOR_HARMON_ALPHA --color_harmon_random --gaussian_options GAUSSIAN_OPTIONS GAUSSIAN_OPTIONS --debug --blending_methods BLENDING_METHODS BLENDING_METHODS --pyramid_blending_levels PYRAMID_BLENDING_LEVELS --distractor_objects DISTRACTOR_OBJECTS DISTRACTOR_OBJECTS

• AUGMENTATION_PATH: Path to a albumentations augmentation file.
• MAX_OBJECTS_PER_IMAGE: The maximum number of objects per image.
• IMAGE_WIDTH: The width of the generated images.
• IMAGE_HEIGHT: The height of the generated images.
• FIXED_IMAGE_SIZES: If set, the images will have the same size.
• SCALE_FOREGROUND_BY_BACKGROUND_SIZE: If set, the foreground objects will be scaled by the background size.
• SCALING_FACTORS: The scaling factors for the foreground objects.
• AVOID_COLLISIONS: If set, the objects will be placed in a way that they do not overlap.
• PARALLELIZE: If set, the generation will be parallelized using multiple processes.
• YOLO_INPUT: If set, the background images can contain yolo annotations.
• YOLO: If set, the generated images will have yolo annotations. Else COCO annotations will be used.
• COLOR_HARMON_ALPHA: The alpha value for the color harmonization.
• COLOR_HARMON_RANDOM: If set, the color harmonization will be random.
• GAUSSIAN_OPTIONS: The gaussian options for the blending. kernel_size and sigma (e.g. 5 1).
• DEBUG: If set, the debug mode will be activated.
• BLENDING_METHODS: The blending methods for the foreground objects. See below.
• PYRAMID_BLENDING_LEVELS: The number of pyramid blending levels.
• DISTRACTOR_OBJECTS: The names of foreground objects which should be used as distractor objects. (Not implemented yet, but will exclude these objects from the annotation file.)

Blending Methods

The blending methods are defined as follows:

• 'ALPHA': Alpha blending.
• 'GAUSSIAN': Gaussian blending.
• 'PYRAMID': Pyramid blending.
• 'POISSON_NORMAL': Poisson blending with normal blending (using the cv2.seamlessClone() function).
• 'POISSON_MIXED': Poisson blending with mixed blending (using the cv2.seamlessClone() function).

The blending methods can be combined with a space.

Mix Datasets

synthetic-data-gen mix --input_dirs INPUT_DIRS --output_dir OUTPUT_DIR --output_splits OUTPUT_SPLITS --percent_sets PERCENT_SETS --test_dataset TEST_DATASET --fixed_data_path FIXED_DATA_PATH --class_names CLASS_NAMES

• INPUT_DIRS: The directories where the datasets are stored.
• OUTPUT_DIR: The directory where the mixed dataset will be stored.
• OUTPUT_SPLITS: The output splits for the mixed dataset. (eg. 0.8 0.2 => 80% train, 20% validation)
• PERCENT_SETS: The percentage of the datasets which should be used. (eg. 0.5 0.5 => 50% of each dataset)
• TEST_DATASET: The dataset which should be used as test dataset.
• FIXED_DATA_PATH: Use a absolute path in the data.yaml file.
• CLASS_NAMES: The class names for the mixed dataset. (eg. class_1 class_2 class_3)

Contribute

1. Clone the repository:

git clone

2. Install the required packages:

pip install -r requirements.txt

3. Install the package in editable mode:

pip install -e .

Publish

1. Update the version in pyproject.toml.

2. Update the CHANGELOG.md.

3. Build the package:

pip install --upgrade build
python -m build

4. Publish the package:

pip install --upgrade twine
python -m twine upload dist/*