yolo-tiling 0.0.7

Tile (slice) YOLO Dataset for Small Objects Detection and Instance Segmentation

YOLO Dataset tiling

This module can cut images and corresponding labels from YOLO dataset into tiles of specified size and create a new dataset based on these tiles. It supports both object detection and instance segmentation. Credit for the original repository goes to slanj.

Installation

To install the package, use pip:

pip install yolo-tiling

Usage

from yolo_tiler import YoloTiler, TileConfig

src = "path/to/dataset"  # Source YOLO dataset directory
dst = "path/to/tiled_dataset"  # Output directory for tiled dataset

config = TileConfig(
    # Size of each tile (width, height). Can be:
    # - Single integer for square tiles: slice_wh=640
    # - Tuple for rectangular tiles: slice_wh=(640, 480)
    slice_wh=(640, 480),

    # Overlap between adjacent tiles. Can be:
    # - Single float (0-1) for uniform overlap percentage: overlap_wh=0.1
    # - Tuple of floats for different overlap in each dimension: overlap_wh=(0.1, 0.1)
    # - Single integer for pixel overlap: overlap_wh=64
    # - Tuple of integers for different pixel overlaps: overlap_wh=(64, 48)
    overlap_wh=(0.1, 0.1),

    # Image file extension to process
    ext=".png",

    # Type of YOLO annotations to process:
    # - "object_detection": Standard YOLO format (class, x, y, width, height)
    # - "instance_segmentation": YOLO segmentation format (class, x1, y1, x2, y2, ...)
    annotation_type="instance_segmentation",

    # For segmentation only: Controls point density along polygon edges
    # Lower values = more points, higher quality but larger files
    densify_factor=0.5,

    # For segmentation only: Controls polygon smoothing
    # Lower values = more details preserved, higher values = smoother shapes
    smoothing_tolerance=0.1,

    # Dataset split ratios (must sum to 1.0)
    train_ratio=0.7,  # Proportion of data for training
    valid_ratio=0.2,  # Proportion of data for validation
    test_ratio=0.1,   # Proportion of data for testing

    # Optional margins to exclude from input images. Can be:
    # - Single float (0-1) for uniform margin percentage: margins=0.1
    # - Tuple of floats for different margins: margins=(0.1, 0.1, 0.1, 0.1)
    # - Single integer for pixel margins: margins=64
    # - Tuple of integers for different pixel margins: margins=(64, 64, 64, 64)
    margins=0.0,

    # Include negative samples (tiles without any instances)
    include_negative_samples=True
)

tiler = YoloTiler(
    source=src,
    target=dst,
    config=config,
    num_viz_samples=15,  # Number of samples to visualize
)

tiler.run()

The tiler requires a YOLO dataset structure in both source and target directories. If only a train folder exists, the train / valid / test ratios will be used to split the tiled train folder; else, the ratios are ignored.

dataset/
├── train/
│   ├── images/
│   └── labels/
├── valid/
│   ├── images/
│   └── labels/
├── test/
│   ├── images/
│   └── labels/
└── data.yaml  # Optional

Command Line Usage

You can also use the command line interface to run the tiling process. Here are the instructions:

python src/yolo_tiler.py --source --target [--slice_wh SLICE_WH SLICE_WH] [--overlap_wh OVERLAP_WH OVERLAP_WH] [--ext EXT] [--annotation_type ANNOTATION_TYPE] [--densify_factor DENSIFY_FACTOR] [--smoothing_tolerance SMOOTHING_TOLERANCE] [--train_ratio TRAIN_RATIO] [--valid_ratio VALID_RATIO] [--test_ratio TEST_RATIO]

Test Data

python tests/test_yolo_tiler.py

Example Commands

1. Basic usage with default parameters:

python src/yolo_tiler.py --source tests/detection --target tests/detection_tiled

2. Custom slice size and overlap:

python src/yolo_tiler.py --source tests/detection --target tests/detection_tiled --slice_wh 640 480 --overlap_wh 0.1 0.1

3. Custom annotation type and image extension:

python src/yolo_tiler.py --source tests/segmentation --target tests/segmentation_tiled --annotation_type instance_segmentation --ext .jpg

Memory Efficiency

The tile_image method now uses rasterio's Window to read and process image tiles directly from the disk, instead of loading the entire image into memory. This makes the tiling process more memory efficient, especially for large images.

---

Disclaimer

This repository is a scientific product and is not official communication of the National Oceanic and Atmospheric Administration, or the United States Department of Commerce. All NOAA GitHub project code is provided on an 'as is' basis and the user assumes responsibility for its use. Any claims against the Department of Commerce or Department of Commerce bureaus stemming from the use of this GitHub project will be governed by all applicable Federal law. Any reference to specific commercial products, processes, or services by service mark, trademark, manufacturer, or otherwise, does not constitute or imply their endorsement, recommendation or favoring by the Department of Commerce. The Department of Commerce seal and logo, or the seal and logo of a DOC bureau, shall not be used in any manner to imply endorsement of any commercial product or activity by DOC or the United States Government.

License

Software code created by U.S. Government employees is not subject to copyright in the United States (17 U.S.C. §105). The United States/Department of Commerce reserve all rights to seek and obtain copyright protection in countries other than the United States for Software authored in its entirety by the Department of Commerce. To this end, the Department of Commerce hereby grants to Recipient a royalty-free, nonexclusive license to use, copy, and create derivative works of the Software outside of the United States.