Coverage and Pose Diversity in Camera Calibration Images¶

Coverage and diversity describe where the checkerboard was observed and how its pose changed. They complement reprojection error by exposing image sets that fit but sample only a narrow part of the camera's field of view.

Image-plane coverage¶

For each accepted view, the audit computes the center of the four outer detected inner corners and normalizes it by image width and height. The default 4 × 3 grid increments the cell containing that center:

coverage_ratio = occupied board-center cells / 12

The report also stores a corner_density_grid, which counts all detected corner observations per grid cell. Board-center coverage is the decision metric; corner density gives a more detailed view of where constraints exist.

What the metric catches¶

A dataset of 20 centered boards can occupy one cell and produce coverage_ratio = 0.0833. More files did not add new image-plane coverage. Moving the target toward the edges and corners generally samples regions where lens distortion is more visible.

Trade-off¶

A center grid is interpretable and resolution-independent, but it loses within-cell detail. A large tilted board can place corners across several cells while its center occupies only one. Use the density grid and annotated previews alongside the ratio.

Physical board area and scale bins¶

Detected points are inner corners, not the physical target edge. The audit fits a homography and projects a boundary one square beyond the outer inner-corner rows and columns. Its area divided by image area is board_area_ratio.

Accepted area values are divided into five bins between the dataset minimum and maximum. occupied_scale_bins counts non-empty bins.

This measures variation within the observed dataset. Three occupied bins do not prove that the absolute near and far distances are suitable. If every view has the same area, one bin is occupied.

Circular rotation range¶

The target's top inner-corner row defines an in-plane angle. Angles wrap at -180°/180°, so ordinary subtraction is wrong near the boundary.

The audit normalizes angles on a circle, finds the largest empty angular gap, and reports:

smallest covering arc = 360° - largest empty gap

Examples:

179° and -179° span about 2°;
a single orientation spans 0°;
0°, 90°, 180°, and -90° require a 270° covering arc.

The default low-pose-diversity warning threshold is 15°.

Signed perspective¶

For the quadrilateral formed by the outer detected inner corners:

horizontal = (top_length - bottom_length) / max(top_length, bottom_length)
vertical   = (left_length - right_length) / max(left_length, right_length)

The signs retain tilt direction. A target tilted left and one tilted right should not collapse into the same pose descriptor. Dataset metrics report the maximum minus minimum for each signed component.

These ratios are projective shape summaries, not recovered 3D target angles.

Capture guidance¶

Aim for:

board centers across the center, edges, and corners;
several target sizes created by changing distance;
in-plane rotations rather than one horizontal orientation;
target tilt in both horizontal and vertical directions;
a fully visible physical checkerboard boundary.

Do not chase a metric by accepting blurred, clipped, or extremely oblique views. All signals must remain usable together.

Configuration¶

The Python API exposes the coverage grid and warning thresholds:

config = AuditConfig(
    pattern=pattern,
    coverage_cols=4,
    coverage_rows=3,
    min_coverage_ratio=0.35,
    min_scale_bins=3,
    min_rotation_range_degrees=15.0,
)

These three diversity thresholds are Python API configuration in version 0.2.2; the CLI exposes their warnings through --fail-on-warning but does not provide flags to change them.

See the Python API reference and the dataset-validation guide.

Documentation version: 0.2.2 · Updated: 2026-07-31