Detect Near-Duplicate Camera Calibration Views

Near-duplicate images add file count without adding much geometric evidence. OpenCV Calibration Audit compares image-derived pose components before calibration and keeps the sharper view when all components are sufficiently close.

Pose representation

Each complete checkerboard detection contributes:

• normalized board-center coordinates;
• physical board area divided by image area;
• in-plane rotation in degrees;
• signed horizontal perspective;
• signed vertical perspective.

The comparison uses normalized measurements so the default thresholds are not tied to one pixel resolution.

Component distances

Two views are duplicates only if every condition passes:

Component	Distance	Default maximum
Center	Euclidean distance in normalized image coordinates	0.03
Scale	Absolute difference of log board areas	0.08
Rotation	Smallest wrapped angular difference	5°
Perspective	Euclidean distance across horizontal and vertical ratios	0.08

The conjunction matters. Similar centers with clearly different target scale, rotation, or tilt remain distinct.

Which image is kept

Candidates are sorted by descending board-region sharpness and then by case-insensitive relative path for deterministic ties. The first view becomes the retained representative. A later match is rejected with:

NEAR_DUPLICATE_POSE
Pose is near-duplicate of sharper view '<path>'.

Its duplicate_of field points to the retained image.

Limits of the method

This is a heuristic comparison of 2D checkerboard geometry. It does not:

• compare raw image hashes or visual content;
• estimate extrinsics first and compare full 3D camera-to-board transforms;
• account for changes elsewhere in the scene;
• prove that two close views have identical calibration information.

Very tight defaults can retain redundant images; loose values can reject views that contain useful differences. Inspect the annotated report when tuning.

Configure duplicate policy in Python

config = AuditConfig(
    pattern=pattern,
    duplicate_center_distance=0.03,
    duplicate_log_area_distance=0.08,
    duplicate_rotation_degrees=5.0,
    duplicate_perspective_distance=0.08,
)

These thresholds are available through the Python API in version 0.2.2. The CLI uses their defaults.

Related documentation

• Coverage and pose diversity
• Decision codes
• Python API configuration

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