Skip to content

Metrics

Every metric in the registry, grouped by tier. Which of them get computed for a dataset is decided by its quantification profile.

Unit kind says how calibration affects the value: length metrics come out in the image's physical unit, area metrics in that unit squared, and ratio / count / colour / intensity metrics are unaffected by it.

Geometry

Computed from the contour alone.

area the enclosed surface perimeter the closed outline max_diameter the two farthest points circularity 4π · area / perimeter² dimensionless, so calibration never changes it 1.00 a perfect circle 0.70 this contour
The four geometry metrics on one contour. The dashed chord is the genuinely farthest pair of points on this outline, and 0.70 is its real circularity — both are computed from the shape drawn, not sketched.
Key Name Unit kind Description
area Area area Enclosed area of the contour polygon.
perimeter Perimeter length Closed perimeter (arc length) of the contour polygon.
circularity Circularity ratio Dimensionless shape descriptor 4·π·area / perimeter² — 1.0 for a circle, lower for elongated or ragged outlines.
max_diameter Max Diameter length Maximum distance between any two points of the contour.

Appearance

Computed from the contour and the image pixels, so these go stale when the underlying imagery or the outline changes.

Key Name Unit kind Description
mean_color_rgb Mean color (RGB) colour Mean red, green and blue channel value (0–255 each) over the contour's interior.
mean_color_lab Mean color (CIELAB) colour Mean CIELAB colour, using OpenCV's 8-bit scaling (L, a, b each in 0–255).
mean_intensity Mean intensity intensity Mean grayscale luminance (0–255), via OpenCV's standard conversion.

Contextual

Computed from the contour and the other contours around it. A contextual value on one object goes stale when a neighbour moves, even if that object's own outline never changed.

Key Name Unit kind Description
nn_distance Nearest-neighbour distance length Euclidean distance from this contour's centroid to the centroid of the nearest other contour sharing the same parent.
mean_knn_distance Mean distance to 3 nearest neighbours length Mean euclidean distance from this contour's centroid to its up-to-3 nearest same-parent sibling centroids (fewer if the group is smaller).

Only-child contours are omitted, not zero

Both contextual metrics need at least one sibling. A contour with no siblings has no meaningful neighbour distance, so it is omitted from the result entirely rather than reported as 0. Do not read a missing value as a distance of zero.

Who counts as a sibling

Contours sharing the same parent. Root-level contours are siblings of every other root-level contour in the image.

k is fixed at 3

mean_knn_distance uses k = 3, and this is not currently configurable.

Relational

Computed from the contour and its children.

Key Name Unit kind Description
n_children Number of children count Number of contours in the same image that name this contour as their parent — how many direct child objects it contains.

Zero is a real value here

Unlike the contextual metrics, n_children is defined for every contour: a leaf object with no children has value 0, which is a meaningful count rather than a missing value. No contour is omitted.

Extending the registry

The registry is extensible — a metric is a class that declares a key, name, description, tier and unit_kind and implements a batch compute. Adding one does not require a schema change, because measurements are stored in a tall table keyed by (contour_id, metric_key) rather than as columns.