Analyze Annotations

Open the report from the project's Report row in the sidebar, or File ▸ Analyze Annotations…. It's split into sections you switch between at the top, each pairing a short explanation with charts. Click a chart bar to see just those images in the list.

The sections

• Overview: image, annotation, and label counts, how many images are empty, and the split breakdown.
• Class Balance: per-label counts and how skewed (or unlabeled) the classes are.
• Geometry: image sizes and aspect ratios (so you can plan a consistent resize), plus box-size and aspect-ratio spreads, outliers, and where objects sit in the frame. A center bias can train models to overweight the middle of the image.
• Validity Issues: the per-annotation findings, grouped by check.
• Export Integrity: problems that might block or corrupt an export.

What it checks

Each finding is Info (often intentional), Warning (probably a mistake), or Error (invalid or export-blocking):

• Boxes: out-of-bounds, zero-area, very tiny, extreme aspect ratios, or heavily-overlapping duplicates.
• Polygons: fewer than three points, an oversized vertex count (usually a machine-traced outline; a Simplify fix is offered), or a saved box that doesn't match the outline (often intentional, flagged for review only).
• Images: no annotations (negatives are fine, and a marked-null ∅ image counts as confirmed), a missing pixel size, a duplicated name, or a near-duplicate of another image caught by perceptual hashing. A near-duplicate in a different split is flagged as data leakage, which inflates every score.
• Labels: annotations with no label, which every export format drops.
• Files (opt-in, Settings ▸ Report ▸ Verify image files): each referenced file checked for missing, unreadable, or corrupt data, with a Remove fix and a Re-link path for moved files.

Fixing issues

Every finding has an Open link that jumps to the image and selects the annotation. Mechanical problems also offer a Fix or Fix all… (with a preview and count):

• Delete duplicate, zero-area, or invalid shapes.
• Clamp to frame: pull an out-of-bounds box or polygon back inside the image.
• Remove near-duplicates: for each group of visually identical images, keep the most-annotated one and remove the rest from the project. Especially useful after a video-frames import at a high frame rate. Files on disk are never deleted.

The rest are review-only: they probably need your judgment, so edit them in the canvas. A fix applied from the report isn't undoable there; for a reversible edit, open the image and change it in the editor.

Issue flags in the list

Images with findings show a severity-colored badge on their thumbnails, List view adds a sortable Issues column, and the toolbar's Filter menu can narrow the list to Has Issues or No Issues. After you edit a project, a banner notes the report is out of date with a Re-analyze button; turn on Settings ▸ Report ▸ Re-analyze automatically after edits to have it recompute shortly after you stop editing.

Tuning (Settings ▸ Report)

The thresholds are yours to adjust: what counts as a tiny box or a rare class, the oversized-polygon vertex limit, and the chart and outlier parameters. Raise the tiny-box threshold for wide-field or sports imagery where small objects are valid. Threshold changes apply on the next analysis, so re-run the report.