Datasets:
File size: 5,876 Bytes
292d868 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 | # Annotation and agreement
Annotators assessed each image's plausibility and naturalness, selected
violation classes, and drew a box around each labelled instance. They could
instead mark no violation or report an unusable image. The released first-pass
records are preserved separately from the resolved final labels.
## The rating rubric
Plausibility asks whether the depicted situation could occur in Singapore.
Naturalness asks whether the pixels resemble a real photograph. An unusual
scenario can be plausible, and a plausible scenario can look visibly generated.
Annotators were instructed to judge these dimensions separately.
| Rating | Plausibility: could this happen here? | Naturalness: does it look real? |
| ------ | ------------------------------------- | ------------------------------- |
| 1 | Impossible | Obvious fake |
| 2 | Unlikely | Suspicious |
| 3 | Possible | Uncertain |
| 4 | Likely | Convincing |
| 5 | Certain | Indistinguishable |
The naturalness guidance highlights visual artefacts such as gibberish text,
extra fingers and warped edges. These are whole-image ratings, not separate
ratings for each violation box.
## Which annotations are compared?
There are 23,710 first-pass submissions for the released 11,534 images: 10,905
images have two raters, 616 have three, and 13 have four. One submission reports
an unusable image and is excluded from eligible comparisons. Its original record
remains in the export; that image has other usable first passes and a usable
final annotation.
For label-overlap and box matching, every eligible rater pair is compared and
pair counts are averaged within each image. More raters do not multiply that
image's pair-count contribution. Exact class-set agreement instead requires all
eligible raters on an image to select the same set. Final gold is not included
as another rater: agreement is measured before the resolution process described
in [ground-truth methods](ground-truth-methods.md).
## Reading the agreement results
| Measure | What it answers |
| ----------------------------------- | ------------------------------------------------------------------------------------------------------- |
| Ordinal Krippendorff's alpha | How consistent are quality ratings after accounting for chance disagreement and their ordering? |
| Exact / within-one rating agreement | How often do all eligible ratings on an image match, or span at most one point? |
| Exact class-set agreement | How often do annotators select precisely the same classes, including agreed negatives? |
| Positive-label agreement | How much do their selected class sets overlap, without giving credit merely for both selecting nothing? |
| Instance F1 | How well do their individual boxes match one to one, with the same class and IoU at least 0.50? |
Positive-label agreement uses twice the number of shared labels divided by the
sum of both label-set sizes. Instance F1 uses twice the number of matched boxes
divided by the total number of boxes from both annotators. The published
estimates weight pair counts equally within each image, pool those weighted
counts across images, then calculate the ratio. They do not average pairwise F1
scores and are not model accuracy scores against gold.
Alpha can be negative when disagreement exceeds its chance baseline. An
undefined statistic is stored as null with its reason rather than interpreted as
zero. Per-class nominal alpha concerns presence/absence of an individual class;
it is not pooled into a single alpha across different classes. Interval alpha
and alternative IoU thresholds are secondary sensitivity measures.
Intervals resample images, preserving all their raters, for 2,000 bootstrap
replicates. The artifact records the fixed seed. Domain label/box agreement uses
all images with only that domain's classes; domain quality agreement selects
images with a final gold label in that domain. Quality-selected subsets overlap.
Support counts and limited-support flags remain in the detailed results.
[Agreement JSON](../evaluation/dataset/annotation_agreement.json) contains
`groups` for the population/domain/class estimates and `bootstrap` for interval
settings. The [CSV](../evaluation/dataset/agreement.csv) provides tabular
results. The
[reproduction guide](evaluation.md#reproduce-the-published-primary-results)
explains how to recalculate them from the released first passes.
## Final quality ratings are a different result
[Image-quality results](../evaluation/dataset/image_quality.json) describe the
resolved ratings: mean plausibility is 3.413 and mean naturalness is 3.372, with
median 4 for both. Its `slices` contains overall, domain and class
distributions. Those descriptive mean intervals use 1,000 image-bootstrap
replicates, unlike the 2,000 used for first-pass agreement.
Resolution can select the minimum human rating or an explicitly adjudicated
rating. Consequently these distributions describe the released annotations; they
are not an average of independent human opinions. Domain/class slices refer to
whole-image ratings and overlap on multi-label images.
The exports do not establish independence from shared prefill. Agreement can
reflect that workflow as well as rater consistency, and neither agreement nor
final resolution proves label correctness. See the
[limitations](limitations.md#annotation-dependence-and-resolved-labels).
|