| # Intended use and limitations |
|
|
| MunicipalBench measures image-level violation classification and instance |
| localisation on a fixed synthetic test set. Use it to compare systems under the |
| same scoring rules and to inspect differences across classes and domains. It has |
| no training split and has not been validated for operational municipal report |
| processing. |
|
|
| ## Synthetic scenes and available context |
|
|
| All distributed images are synthetic. A high score therefore does not establish |
| accuracy on real reports, where capture conditions, scene frequencies and |
| ambiguity can differ. Historical real controls are omitted and do not provide a |
| released real-world validation set. |
|
|
| Some classes require information outside the image, such as whether a vehicle |
| has permission to park. The taxonomy records definitions and exclusions, but a |
| single frame cannot verify every condition. Treat scores as agreement with the |
| benchmark's visual labels, not legal determinations. |
|
|
| Generation artefacts and repeated templates may influence predictions. The |
| release does not establish independence between images from the same generation |
| family. Its image-bootstrap intervals do not account for that unrecorded |
| clustering. Publicly available images and labels also create a contamination |
| risk; disclose prior exposure and record the dataset revision. |
|
|
| ## Annotation dependence and resolved labels |
|
|
| First-pass records do not rule out shared prefill. High agreement, particularly |
| box overlap, may partly reflect the annotation interface rather than independent |
| judgements. Conversely, low rating agreement indicates that raters did not |
| consistently apply those quality judgements; it is not repaired by evaluating |
| agreement after resolution. |
|
|
| Final labels combine automatic rules, VLM votes, human adjudication and earlier |
| manager selections. These routes have different guarantees. For example, the |
| native automatic box criterion can pass despite unmatched instances, and VLM |
| votes can retain shared errors. The release records the method for every image |
| so those differences remain inspectable. It does not claim an independent expert |
| correctness audit of all final labels. |
|
|
| Resolved quality ratings can use conservative minima. Their means and histograms |
| describe final labels rather than an unbiased survey of image quality. See the |
| [annotation protocol](annotation-protocol.md) and |
| [resolution rules](ground-truth-methods.md). |
|
|
| ## Uneven support and comparison limits |
|
|
| Class frequencies vary, and multi-label images belong to several domain slices. |
| Micro scores give more influence to frequent targets; macro scores and per-class |
| results expose a different view. Small differences between model point estimates |
| are not evidence of statistically significant superiority. Model-score tables do |
| not include paired confidence intervals for differences. |
|
|
| The model comparison has specific execution limitations, including Gemini's |
| mixed box-axis interpretation and mutable hosted providers. These remain in |
| [model provenance](../evaluation/models/synthetic11534/README.md). The reported |
| scores describe saved outputs under those settings; they are not controlled |
| hardware, latency or cost comparisons. |
|
|
| Semantic relaxation can credit related but genuinely distinct classes, so its |
| higher scores must not replace exact-class results. Historical studies use old |
| gold, different populations and, in some cases, only three runs or a |
| positive-only subset. Their individual limits are explained beside their |
| findings in the [supplementary guide](supplementary.md). |
|
|
