Crop Image

1 VersionTransform

Crops an Image to a region given as a BoundingBox, Rectangle<Double>, Polygon<Double>, Mask, or Ellipse<Double>, returning either a tight crop or the original-size frame with outside pixels painted by a configurable fill colour.

How it fits

Typical backends

A detector locates objects, the crop isolates each region, and OCR runs only on the cropped frame so text recognition is not distracted by background.

A segmentation mask drives the crop, which paints everything outside the segmented region; the masked frame is then captioned for a focused description.

Caveats

FallbackA shape that cannot be rasterised — a Mask whose height / width disagree with the image, a Polygon<Double> with fewer than three points, an Ellipse<Double> with a non-positive radius, or an all-zero mask — is logged and treated as an empty region, emitting a `1x1` fill-coloured pixel instead of failing the stream; downstream must detect and discard that placeholder.
I/O contractThe fill is read as a `(B, G, R)` triple via background; supplying a BGR fill for an RGB-tagged image swaps red and blue with no warning.
Parameter interactionWith keep_size off (the default) the output is sized to the region's bounding rectangle and varies tick to tick; turning it on returns the original frame size with outside pixels painted by the fill colour.
AccuracyPolygon<Double> vertices are truncated to integers (not rounded) before rasterisation, so sub-pixel polygons lose precision at the region boundary.
I/O contractThe output color space matches the input — there is no colour conversion — so pair with convert_image_format when a downstream stage needs a specific layout.
State lifetimeThe fill colour and sizing mode are captured once at startup; changing them requires a restart.

Versions

a10310e2latestdefault
Automated release
6/25/2026