COT5415_Computer_Vision / docs /data /QUALITY_POLICY.md
iiCEMAN's picture
Add PACO-LVIS (v2 unified release, annotations only; images from COCO 2017)
b0d869f verified
|
Raw History Blame Contribute Delete
3.52 kB

Source geometry and background policy

Converted records retain every original annotation and raw polygon. Non-decodable geometry receives mask_status and an empty placeholder; partially decodable geometry retains its valid raster components and raw geometry but is not exposed as a usable part mask by the shared loader. Never replace an invalid polygon with a guessed box or silently drop its record.

Across PartImageNet: 168 invalid-geometry annotations (train 121, val 17, test 30). A separate valid-but-empty source mask is counted in the test conversion summary; it is not an additional malformed polygon. Whole masks are derived unions, so empty source content may produce empty whole nodes too. These categories must not be conflated.

Default loader: exclude mask_status other than valid from sample['masks']; retain the node in sample['record'] and list it in invalid_node_ids. A valid empty mask remains represented as source data. Training consumers must explicitly define any further filtering or loss mask; this package does not silently infer a supervision objective.

Current PartImageNet validation experiment: exclude the 17 invalid masks, retain all 2,957 images, and keep uncovered pixels at background. This affects 16 images. There is no ignore mask or image-level exclusion. This choice follows the released loader convention and is recorded with every affected image; it is not a claim that malformed regions were recovered.

Pascal source PNGs retain all values. Verified named metadata is a sidecar lookup, not a rewrite of masks, category keys or splits. Background IDs and observed-label scope are documented with the source-backed label table. No new ignore ID is manufactured for absent labels.

Validation distinguishes conversion integrity from source quality. The final report has conversion_integrity_passed=true and zero conversion issues; strict passed=false reflects preserved source defects. A nonzero strict validator exit in this state is expected and must not be hidden or changed to pass.

Source-backed Pascal metadata now confirms whole background=16 and part background=54, with no observed ignore code. A semantic-parent diagnostic flags 871,870 train and 706,599 val foreground-part pixels whose expected whole class differs from the supplied whole PNG. It affects 512 train and 547 val images. This diagnostic does not rewrite the original rasters or their lossless converted intersections. A consumer defining semantic targets should explicitly exclude or separately handle these nodes. Part-background pixels inside a foreground whole are recorded separately and are not counted as a foreground semantic-class conflict.

PACO-LVIS (added 2026-11-18). All object polygons in the three official splits rasterize as valid geometry. 123 masks are empty after decoding (train 108, val 3, test 12); they are kept and counted, not removed. PACO part masks and LVIS object polygons were drawn independently, so most parts (79%) extend at least one pixel beyond their parent object's mask. The part masks are copied verbatim and not clipped. Each part's outside_parent_area records the overflow, and validation recomputes it. Consumers who need strict containment should intersect a part with its parent explicitly and record that choice. PACO's federated LVIS labels also matter: a category absent from an image is only a confirmed negative if it is listed in source_image_metadata.neg_category_ids; categories in not_exhaustive_category_ids may have unlabelled instances.