PD-Mera/object-detection-data-synthesis

Synthesis data in YOLO format given background and object images

Object Detection Data Synthesis

Data Synthesis pipeline to generate object detection data

How to run

Run with pip

pip install mdetsyn

And run in python file

from mdetsyn import run_synthesis, create_args

args = create_args()
run_synthesis(args)

Run with command line

python synthesis.py --backgrounds ./backgrounds \
                    --objects ./objects \
                    --savename ./synthesis \
                    --number 1000 \
                    --class_mapping ./class_mapping.json \
                    --class_txt ./classes.txt

Sample

• Backgrounds folder contain background images (in any folmat)

├── backgrounds/
    ├── background-0.jpg
    ├── background-1.jpg
    └── ...

• Objects folder contain object images in subfolders (the best is .png format with A channel but any format is still runnable)

├── objects/
    ├── class_1/
    │   ├── image-0.png
    │   ├── image-1.png
    │   └── ...
    ├── class_2/
    └── ...

• Each image in objects folder will be synthesis by n times with n is user input

• Output is a synthesis folder contain images and labels dir same as YOLO format

• Sample visualization:

Background	Object	Synthesis

Support synthesis methods

• Random Resize
• Random Rotate
• Random Transparency
• Random Perspective Transform
• Seamless Clone
• Grayscale

Error and TODO

• Sometimes seamless clone does not work
• Input parameter for each augment
• Add default arguments to argparse help