https://github.com/pd-mera/object-detection-data-synthesis

Synthesis data in YOLO format given background and object images
https://github.com/pd-mera/object-detection-data-synthesis

data-synthesis yolo

Last synced: 3 months ago
JSON representation

Synthesis data in YOLO format given background and object images

• Host: GitHub
• URL: https://github.com/pd-mera/object-detection-data-synthesis
• Owner: PD-Mera
• Created: 2024-04-24T06:52:37.000Z (about 1 year ago)
• Default Branch: master
• Last Pushed: 2024-05-08T10:24:33.000Z (about 1 year ago)
• Last Synced: 2025-01-08T19:46:54.190Z (5 months ago)
• Topics: data-synthesis, yolo
• Language: Python
• Homepage:
• Size: 485 KB
• Stars: 2
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# Object Detection Data Synthesis

Data Synthesis pipeline to generate object detection data



## How to run

### Run with pip

``` bash

pip install mdetsyn

```

And run in python file 

``` python

from mdetsyn import run_synthesis, create_args

args = create_args()

run_synthesis(args)

```

### Run with command line

``` bash

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 |

| :---: | :---: | :---: |

| ![](./sample/backgrounds/background-0.jpg) | ![](./sample/objects/vanamo/vanamo-0.png) | ![](./sample/synthesis/images/4b09b7ae-3290-45d7-afef-82a7c5cef26d.jpg) |

## Support synthesis methods

- Random Resize

- Random Rotate

- Random Transparency

- Random Perspective Transform

- Seamless Clone

- Grayscale

## Error and TODO

- [ ] Sometimes seamless clone does not work

- [x] Input parameter for each augment

- [ ] Add default arguments to argparse help