geotiff-cutout 0.1.1

Create rectangular pixel cutouts from GeoTIFFs centred on GeoJSON features, with COCO annotation output.

geotiff-cutout

Create rectangular pixel cutouts from a GeoTIFF centred on GeoJSON Polygons or Points, with an optional random shift. All intersecting annotations are returned in COCO format — boxes that extend beyond the cutout edge are clipped and flagged with is_truncated, Polygons will also be returned as segmentation. Classes must be indicated in the properties of the features.

---

Installation

pip install geotiff-cutout

Note: rasterio and fiona depend on GDAL. On most platforms the wheels ship GDAL bundled, so a plain pip install is sufficient. If you hit build errors, install via conda:

conda install -c conda-forge rasterio fiona
pip install geotiff-cutout

---

Quick start

Python API

from geotiff_cutout import make_cutout

# Rectangular cutouts with moderate random shift saved to disk
make_cutouts(
    "image.tif",
    "annotation.geojson",
    "data",
    coco_path="annotations.json",
    cutout_width_px=400,
    cutout_height_px=200,
    random_shift=0.5,
    method="fallback"
)

Command line

# Square cutouts, default settings
geotiff-cutout ortho.tif annotations.geojson

# Rectangular cutouts with custom shift and output directory
geotiff-cutout ortho.tif annotations.geojson \
    --width 640 --height 480 \
    --shift 0.8 \
    --output my_cutouts/

# Full help
geotiff-cutout --help

---

The random_shift parameter

Value	Effect
0.0	Cutout centred exactly on the feature centroid.
0.5	Centroid can wander up to half the available slack.
1.0	Maximum shift: the target bbox may touch (but never cross) the cutout edge.

Shift is applied independently on each axis, so rectangular cutouts behave correctly — a wide cutout allows more absolute horizontal shift than a tall one.

---

COCO output format

Each call returns a standard COCO dict:

{
  "images": [{
    "id": 0
    "width": 512,
    "height": 512,
    "file_name": "cutout_<uuid>.png",
    "geotiff_source": "ortho.tif",
    "geojson_source": "annotations.geojson",
    "cutout_col_offset": 340,
    "cutout_row_offset": 120
  }],
  "categories": [{"id": 1, "name": "car"}],
  "annotations": [{
    "id": 0,
    "image_id": 0,
    "category_id": 1,
    "bbox": [12.0, 34.5, 80.0, 60.0],
    "bbox_mode": "xywh",
    "area": 4800.0,
    "iscrowd": 0,
    "is_truncated": false,
    "original_bbox": [12.0, 34.5, 80.0, 60.0],
    "segmentations":[[12.0, 34.5, ...,  80.0, 60.0]],
    "origina_segmentations":[[12.0, 34.5, ...,  80.0, 60.0]]
  }]
}

all coordinates in coco annotation are local image pixel coordinates bbox is the annotation clipped to the cutout boundary ([x, y, w, h] in local cutout pixels).
original_bbox is the full, unclipped box (may have negative coordinates or values exceeding the cutout dimensions for truncated boxes).
segmentation similar to bbox but Polygon segemnt if input was not Point. original_segmentation similar to original_bbox but Polygon segemnt if input was not Point. is_truncated is true when the box was clipped.

The category name is read from the first matching GeoJSON property key: category → class → label → "object" (fallback).

---

License

AGPL 3