On this page
Define the mask contract before looking at pixels
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. Record a controlled image id, image width and height, mask width and height, channels or planes, storage type, and the transform from image to mask. Then identify whether values are class ids, instance ids, binary foreground values, or a color palette.
| Contract field | Synthetic example | Review question |
|---|---|---|
| Image geometry | 640 by 480 pixels | Which image version is the mask paired with? |
| Mask geometry | 640 by 480 pixels | Does the mask align without a crop or resize? |
| Encoding | integer class id | Could a palette or alpha channel change meaning? |
| Background | 0 means background | Is 0 a class, background, or unknown? |
| Label map | 1=road, 2=tree | Which version defines every nonzero value? |
| Instance policy | class mask only | Are separate objects intentionally merged? |
| Transform | none declared | Was the mask rotated, padded, or resized? |
The cited segmentation-mask documentation is a format reference, not a source of domain labels. A file that opens and has two dimensions can still be semantically wrong if a palette, value range, or label-map version is missing. Keep those findings distinct.
Review class ids and background semantics explicitly
A label map is part of the annotation contract. Record the numeric or color value, stable label id, display label, definition version, and whether the value is allowed in this task. Keep background, ignore, unknown, and unlabeled states separate when the source distinguishes them. Do not turn an absent label map into a guessed class name.
Synthetic label-map and mask summary
{
"image_id": "synthetic-frame-02",
"mask": {"width": 4, "height": 3, "encoding": "uint8", "channels": 1},
"label_map_version": "synthetic-v2",
"values": [
{"value": 0, "state": "background"},
{"value": 1, "label_id": "synthetic:road", "state": "class"},
{"value": 2, "label_id": "synthetic:tree", "state": "class"},
{"value": 255, "state": "ignore"}
],
"observed_values": [0, 1, 2, 255],
"disposition": "conditional"
}| Observation | Safe interpretation | Do not infer |
|---|---|---|
| Unexpected value | The value is outside the supplied map | Which class it should become |
| No pixels for a mapped class | The class was not observed in this fixture | That the class is absent everywhere |
| 255 appears | An ignore or special state may be present | That it means background |
| Palette differs | Encoding metadata changed or is incomplete | That displayed colors are ids |
| All values fit map | Observed values have definitions | That annotations are accurate |
Reconcile mask shape with image and transform history
Compare image and mask dimensions, origin, axis direction, and transform order. A one-pixel shift can make an otherwise plausible mask overlap the wrong objects. A resize can change boundaries; a crop can remove context; a rotation can change orientation. Preserve source geometry and derived comparisons rather than overwriting the mask.
| Check | Evidence to record | Hold condition |
|---|---|---|
| Dimensions | Image and mask width and height | One dimension is unknown |
| Pixel origin | Top-left or declared alternative | Origin is absent |
| Transform chain | Crop, pad, resize, rotate, mirror | Only final dimensions are known |
| Alignment sample | Synthetic landmarks or approved visual note | No alignment evidence is available |
| Storage type | Bit depth, channels, palette or integer map | Decoder output is treated as source truth |
| Instance policy | Class-only, instance ids, or merged objects | Instance meaning is guessed |
The CVAT Documentation, (opens in a new tab) documentation distinguishes segmentation representations such as polygons and run-length encodings. A polygon, an RLE mask, and a raster mask can describe related regions while preserving different precision and boundary behavior. Record the representation used at each stage and any rasterization rule.
Use a synthetic failure matrix before source review
Use invented four-by-three fixtures to exercise the review. Include an offset mask, a missing label-map entry, a value that means ignore, a palette mismatch, and a class-only mask where instance separation is unavailable. The goal is to prove that the worksheet keeps facts and decisions separate, not to score a segmentation model.
| Fixture | Observed fact | Disposition |
|---|---|---|
| same-shape-aligned | Image and mask dimensions match; map resolves values | Conditional, alignment evidence still needed |
| one-column-shift | Dimensions match but synthetic landmark is displaced | Hold transform record |
| unknown-value | Mask contains an id absent from the map | Hold label-map owner decision |
| ignore-value | Special value is documented as ignore | Keep separate from background |
| palette-change | Rendered color changed while ids are unavailable | Hold encoding evidence |
| merged-instances | One class id covers multiple regions | Record class-only instance policy |
Keep mask evidence scoped and qualified
The arXiv / Cornell University, (opens in a new tab) provides dataset context, and the cited format references describe interchange fields. Neither source proves that a particular mask is correct, complete, authorized, private, or suitable for a downstream use. A review can document dimensions, encoding, label-map coverage, transform evidence, and unresolved ambiguity.
Pair this guide with annotation label ontology versioning for label identity and bounding-box coordinate conventions for region geometry. 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 decode masks, convert labels, create a quote, match a buyer, issue a legal opinion, or grant a license.
Takeaways
- Record image and mask geometry, encoding, background, label map, instance policy, and transforms separately.
- Treat unknown and ignore values as explicit states, not guessed classes.
- Use synthetic landmarks and small matrices to expose shifts and palette mistakes.
- Keep polygons, RLE, and raster masks traceable as different representations.
- Separate mask evidence from annotation accuracy, privacy, rights, and intended use.