On this page
Write the coordinate contract before comparing boxes
The source dates shown in this guide identify documentation checks on 22 October 2026, not original publication dates. Follow each citation for the format vocabulary being discussed. First record a controlled image id, width, height, frame or page index, and any resize or crop operation. Then record whether coordinates are pixels, normalized fractions, or another declared unit; whether the origin is the top-left or another point; and whether the fourth fields mean width and height or maximum x and y.
| Contract field | Synthetic example | Question to resolve |
|---|---|---|
| Image geometry | width 640, height 480 | Which dimensions belong to this annotation? |
| Origin | top-left | Does y increase down the image? |
| Units | pixel coordinates | Are values absolute or normalized? |
| Field order | x, y, width, height | Could the last pair be x_max, y_max? |
| Edge convention | half-open or inclusive | How is a one-pixel edge represented? |
| Transform | crop then resize | Which transform produced these numbers? |
| Clip policy | inside image or flagged | Are outside values errors or evidence? |
The cited format documentation shows a format with image dimensions and object annotation fields such as bbox, while segmentation is carried separately CVAT Documentation, (opens in a new tab). Do not infer every field convention from a familiar name. Preserve the declared format and the local export documentation as evidence.
Reconcile boxes with image bounds and transforms
For a synthetic image with width W and height H, a pixel box in x, y, width, height form has a start point and an extent. A review can derive an end point for comparison, but it should not silently rewrite the source. Keep both the raw values and the derived comparison values, with the convention used for the derivation.
Synthetic box worksheet fixture
{
"image": {"id": "synthetic-frame-01", "width": 640, "height": 480},
"source_box": {"format": "xywh", "units": "pixel", "origin": "top-left", "x": 120, "y": 80, "width": 160, "height": 120},
"derived": {"x_end": 280, "y_end": 200, "inside_image": true},
"review": {"transform": "none declared", "edge_rule": "unknown", "disposition": "conditional"}
}The fixture is invented and contains no customer image or private label. If a source uses normalized values, record the image dimensions used at conversion time. A box that fits after a resize can still be wrong if it was measured against the pre-crop geometry. Record crop, pad, rotate, and mirror operations in order, rather than keeping only a final guessed box.
| Finding | What it establishes | What it does not establish |
|---|---|---|
| Negative x or y | The box is outside or uses another convention | That clipping is safe or intended |
| End point beyond bounds | Extent and image geometry disagree | That the object is absent |
| Width or height zero | The supplied extent is not a visible area | Why the annotation was emitted |
| Fractional pixel | The source uses subpixel or normalized precision | That rounding is harmless |
| Box fits after transform | One stated transform reconciles values | That the source transform is complete |
Compare formats without losing source meaning
A conversion is a new evidence record. Capture source format, destination format, source image geometry, transform, rounding rule, clipping rule, and the reason for conversion. A xywh to xmax,ymax derivation can be reversible only if the edge convention and numeric precision are known. A crop or rotation may require an explicit geometry transform, not a field rename.
| Conversion question | Record | Hold when |
|---|---|---|
| Field semantics | Source and target names with definitions | Names are guessed from position |
| Image dimensions | Dimensions before and after transform | One side is absent |
| Precision | Decimal scale and rounding operation | Rounding rule is unknown |
| Clipping | Original and clipped coordinates | Out-of-bounds values are discarded |
| Identity | Source annotation id and destination id | The output cannot be traced |
| Validation | Synthetic round-trip or visual review note | Only a parser success is available |
- Read the format declaration and write its field order in the worksheet.
- Recompute a small synthetic fixture in both conventions.
- Compare derived bounds while retaining the unmodified source values.
- Mark clipping, rounding, and unknown edge rules as findings.
- Attach each conversion result to the source artifact and transform record.
Use a box review matrix for disagreements
When two annotators or exports disagree, compare the contract before comparing the object. Keep one row per annotation id and image version. The matrix below is a starting point for a synthetic review; it deliberately avoids deciding whether an object should have been labeled.
| Check | Evidence | Disposition |
|---|---|---|
| Same image id and dimensions | Source id, width, height | Compare geometry or hold identity |
| Same field convention | xywh, xyxy, normalized, or unknown | Convert only with a stated rule |
| Same transform chain | Crop, resize, rotate, pad sequence | Reconcile or record a transform gap |
| Same edge rule | Inclusive, half-open, or unknown | Flag boundary uncertainty |
| Same class identity | Ontology version and label id | Route label issue separately |
| Same region extent | Derived bounds and delta summary | Human review if difference matters |
| Source retained | Raw coordinates plus derived record | Do not overwrite source evidence |
Keep coordinate evidence scoped and qualified
The arXiv / Cornell University, (opens in a new tab) provides context for object localization and segmentation datasets, but a format reference does not prove that any particular box is accurate or suitable for a use. A coordinate review can establish what was declared, how a derived value was computed, and which ambiguity remains. It cannot establish ownership, permission, privacy, annotation quality, or model performance.
Pair this guide with annotation label ontology versioning for class identity and the data provenance log for image and transform history. After supported non-health scope, authority to describe the collection, and a controlled annotation packet are established, the qualified appraisal request path may be a next preparation step. It does not convert boxes, inspect images, create a quote, match a buyer, issue a legal opinion, or grant a license.
Takeaways
- Record geometry, origin, units, field order, edge rule, and transforms before comparing coordinates.
- Keep raw source boxes separate from derived or converted values.
- Use synthetic round trips to expose field-order and rounding mistakes.
- Treat clipping, zero extents, and unknown conventions as explicit findings.
- Separate geometry evidence from label meaning, accuracy, privacy, rights, and intended use.