Python image tiling library for image processing, object detection, etc.
plakakia is an efficient image tiling tool designed to handle bounding boxes within images. It divides images into rectangular tiles based on specified parameters, seamlessly handling overlapping tiles. The tool assigns bounding boxes to tiles that fully contain them, and it also offers an option to eliminate duplicate bounding boxes. While the current version only supports fully contained bounding boxes, future updates will include support for partial overlap. plakakia can handle object detection and segmentation datasets.
Currently, the library offers online and offline modes for processing data (refer to the Usage section section below for more details):
- In the offline mode, one can use a config file and run a script once to process all data.
- In the online mode, the
tile_imagefunction allows processing of images of any dimension.
There are plans to expand plakakia's capabilities in the offline mode to handle images with more than 3 channels.
To ensure optimal performance, plakakia utilizes the multiprocessing and numpy libraries. This enables efficient processing of thousands of images without the use of nested for-loops commonly used in tiling tasks. For detailed benchmarks on various public datasets, please refer to the information provided below.
It is highly recommended that you create a new virtual environment for the installation:
- Download and install Mamba (or Anaconda).
- Create a virtual environment:
mamba create -n plakakia jupyterlab nb_conda_kernels ipykernel ipywidgets pip -y - Activate the environment:
mamba activate plakakia - Run the following command to install the library:
pip install plakakia
make_tiles --config path/to/config.yaml
Here's an example config file.
from plakakia.utils_tiling import tile_image
tiles, coordinates = tile_image(img, tile_size=100, step_size=100)
For more examples, check the examples folder.
You can run the demo app with the following command:
streamlit run demo/explore_tiling_output.py
And when you open http://localhost:8501 in your browser, you should see the following:
Benchmarked on HP Laptop with specs: AMD Ryzen 5 PRO 6650U; 6 cores; 12 threads; 2.9 GHz
| Dataset | Source | Formats (images/labels) | Number of images | tile_size | step_size | tiles generated | plakakia performance |
|---|---|---|---|---|---|---|---|
| Solar Panels v2 | RoboFlow | jpg/COCO | 112 | 150 | 50 | 3.075 | 1,11 sec |
| Traffic Signs | Kaggle | jpg/YOLO | 741 | 300 | 200 | 1.695 | 2,8 sec |
| Hard Hat Workers v2 | RoboFlow | jpg/YOLO | 5.269 | 100 | 50 | 21.678 | 6,94 sec |
| Microsoft COCO dataset | RoboFlow | jpg/YOLO | 121.408 | 200 | 150 | 177.039 | 3 min 4 sec |
☑️ Fix reading of classes from annotations (create a 'mapper' dictionary to map classes to numerical values).
☑️ Read settings from a file (e.g. json).
☑️ Removing all tiles with duplicate bounding boxes (that appear in other tiles).
☑️ Support other annotation formats (e.g. coco). (only input for now)
☑️ Provide tiling functionality without any labels needed.
☑️ Add support for segmentation tasks (tile both input images and masks).
☑️ Add a demo app for the users to be able to see the tiling applied on an image.
⬜️ Add less strict (flexible) duplicate removal methods to avoid missing bounding boxes.
⬜️ Consider bounding boxes in tiles if they partially belong to one.
⬜️ Support reading annotations from a dataframe/csv file.
⬜️ Make tiles with multidimensional data offline with config file (e.g. hdf5 hyperspectral images).
If you want to contribute to this project, please check the CONTRIBUTING.md file.