govtech/municipalbench-full

MunicipalBench

MunicipalBench is an evaluation benchmark of 11,534 synthetic municipal-scene images, with 97 violation classes across 13 domains. It supports image-level multi-label classification and instance localisation. It is not validated as training data or as a measure of performance on real municipal reports.

Version v1.0.0-rc1.

Licence

MunicipalBench is provided under the GovTech Singapore licence, which incorporates MIT terms with additional GovTech conditions. It covers the repository contents subject to the exclusions stated in the licence, including specified government branding, third-party material and restricted assets. The full terms include Singapore governing law and SIAC arbitration.

Contents

Population	Images
Ordinary synthetic	10,838
Synthetic traps	696
Violation present	8,455
No violation	3,079
Total	11,534

All distributed images have final annotations and original first-pass records. There are 19,337 final violation boxes and 23,710 first-pass submissions. Real controls and unusable images are excluded. trap is a source category; it does not imply a negative final annotation.

data/test/                  original images and ImageFolder metadata
annotations/                final annotations, first passes, COCO export
taxonomy/                   class definitions, domains, semantic policy
evaluation/dataset/         composition, IAA and image quality
evaluation/models/          released-population results and predictions
evaluation/supplementary/   historical results and selected secondary analyses
code/                       scoring and IAA source with checksum inventory
examples/                   loading, overlay and reproduction commands
release/                    manifest, provenance and checksums
docs/                       annotation, scoring, supplementary studies and limits

Start with ground_truth.json for complete final annotations. first_pass_annotations.json contains the original human submissions used for inter-annotator agreement (IAA), not a second final ground truth. The optional ground_truth_coco.json is the same labels and boxes in the Common Objects in Context (COCO) interchange format. It is not additional annotation. Native boxes are normalised XYWH; COCO boxes are pixel XYWH. See the data dictionary.

Load and inspect

from datasets import load_dataset

# Pin revision to a reviewed commit SHA for a reproducible evaluation.
dataset = load_dataset("govtech/municipalbench-full", split="test")
print(dataset[0]["image_id"], dataset[0]["objects"])
dataset[0]["image"].show()

After downloading a snapshot, install code/requirements.txt in your Python environment. Run python examples/load_dataset.py --root . to check the ImageFolder dataset, or python examples/render_annotation.py --root . for an annotated preview. No custom remote dataset script is required.

Dataset evaluation

Composition, annotation reliability and final quality ratings answer different questions. Composition results describe class/domain support and label resolution. The agreement results below measure consistency between first-pass annotators. Final quality distributions describe the resolved plausibility and naturalness ratings; the annotation guide explains their scales and why their averages must not be confused with agreement.

The primary IAA analysis uses first passes for exactly the released 11,534 images, before final resolution. One reported submission is excluded from eligible annotation comparisons; its original record remains available.

Measure	Estimate	95% interval
Plausibility ordinal alpha	0.2957	0.2792–0.3119
Naturalness ordinal alpha	0.2242	0.2069–0.2418
Exact class-set agreement	0.5829	0.5739–0.5912
Positive label agreement	0.7248	0.7187–0.7308
Instance F1 at IoU 0.50	0.5726	0.5629–0.5824

Annotation agreement by domain

Domain	Positive-label agreement (95% CI)	Instance F1 (95% CI)
Estate maintenance	0.6623 (0.6524–0.6716)	0.4861 (0.4711–0.4995)
Roads and footpaths	0.6622 (0.6421–0.6816)	0.5081 (0.4849–0.5322)
Animals	0.9235 (0.9035–0.9419)	0.8510 (0.8046–0.8920)
Pests	0.8862 (0.8649–0.9076)	0.7357 (0.6867–0.7838)
Parks and greenery	0.6793 (0.6492–0.7068)	0.5752 (0.5391–0.6082)
Abandoned equipment	0.8252 (0.8019–0.8476)	0.7151 (0.6776–0.7483)
Illegal parking	0.6618 (0.6211–0.7000)	0.6083 (0.5638–0.6519)
Drains and sewers	0.8098 (0.7848–0.8339)	0.7247 (0.6884–0.7582)
Water bodies	0.8197 (0.7953–0.8427)	0.6423 (0.5899–0.6946)
Fire safety and emergency access	0.8058 (0.7741–0.8384)	0.7104 (0.6660–0.7519)
Safety	0.7796 (0.7561–0.8022)	0.7206 (0.6955–0.7469)
Public transport infrastructure	0.8959 (0.8665–0.9210)	0.6688 (0.6176–0.7197)
Street furniture and amenities	0.8906 (0.8636–0.9153)	0.7994 (0.7576–0.8392)

Each domain is evaluated across the 11,534 released images, considering only that domain's labels and boxes. Instance F1 uses exact classes and IoU ≥ 0.50. Parentheses contain the existing 95% image-bootstrap intervals. N/A means undefined; † flags limited support. See the full agreement results and CSV for class-level results, rating agreement, support flags and sensitivity analyses.

Plausibility and naturalness agreement by domain

Domain	Plausibility ordinal alpha (95% CI)	Naturalness ordinal alpha (95% CI)
Estate maintenance	0.2129 (0.1883–0.2406)	0.1938 (0.1664–0.2239)
Roads and footpaths	0.1818 (0.1325–0.2303)	0.1873 (0.1376–0.2369)
Animals	0.1876 (0.0835–0.2861)	0.2485 (0.1439–0.3363)
Pests	0.0533 (-0.0309–0.1279)	0.3809 (0.3092–0.4497)
Parks and greenery	0.1535 (0.0825–0.2227)	0.1370 (0.0650–0.2054)
Abandoned equipment	0.0683 (-0.0131–0.1457)	0.1682 (0.0853–0.2459)
Illegal parking	0.0365 (-0.0688–0.1378)	0.1899 (0.0885–0.2926)
Drains and sewers	0.1038 (0.0204–0.1869)	0.2171 (0.1352–0.2982)
Water bodies	0.1833 (0.1118–0.2530)	0.2912 (0.2201–0.3587)
Fire safety and emergency access	0.2516 (0.1449–0.3504)	0.2333 (0.1209–0.3471)
Safety	0.0588 (-0.0210–0.1318)	0.2139 (0.1386–0.2878)
Public transport infrastructure	0.0010 (-0.0992–0.1094)	0.1314 (0.0220–0.2358)
Street furniture and amenities	0.1839 (0.0671–0.2962)	0.1100 (0.0022–0.2125)

These are ordinal Krippendorff's alpha estimates from first-pass ratings, measuring annotator consistency rather than average image quality. Each domain uses images with at least one final gold label in that domain and eligible rating pairs. These subsets overlap and differ in size. Parentheses contain 95% image-bootstrap intervals; N/A means undefined and † flags limited support. Exact support counts remain in the full agreement results.

Intervals are 95% percentile bootstrap intervals resampling images, with 2,000 replicates. Exact class-set agreement includes agreed negatives; positive agreement and instance F1 provide complementary views. Agreement on quality ratings uses ordinal Krippendorff's alpha. IAA measures consistency, not final-label correctness. Shared prefill may affect independence; generation-family dependence is not modelled. Read the annotation protocol and limitations before interpreting these values.

Final resolution uses GT-001/002/003/004/005/008, with method provenance on each image. Ground-truth methods documents the rules and counts. Agreement is measured on original first passes before resolution.

Model evaluation

The six-model comparison includes YOLO-World, Gemma 4 12B, GLM 5.3 Flash, GPT-5.6 Luna, Muse Glimmer 30B and Gemini 3.8 Flash. Each run covers all 11,534 images with valid responses. The table uses 11,534 shared-valid images, exact class matching and localisation IoU ≥ 0.50.

Model	Classification micro F2	Localisation micro F2
YOLO-World	0.1193	0.0604
Gemma 4 12B	0.3764	0.1454
GLM 5.3 Flash	0.6264	0.3442
GPT-5.6 Luna	0.5937	0.2928
Muse Glimmer 30B	0.5990	0.2676
Gemini 3.8 Flash	0.6710	0.3239

Classification performance by domain

Classification micro F2 for each domain, using exact class matching on the same 11,534 shared-valid images. Counts are pooled across the domain's classes.

Domain	YOLO-World	Gemma 4 12B	GLM 5.3 Flash	GPT-5.6 Luna	Muse Glimmer 30B	Gemini 3.8 Flash
Estate maintenance	0.0633	0.3967	0.5929	0.5659	0.5676	0.6204
Roads and footpaths	0.0759	0.3510	0.6323	0.6638	0.6070	0.6934
Animals	0.3184	0.5131	0.8822	0.8507	0.8784	0.9167
Pests	0.0082	0.2665	0.6872	0.5261	0.7003	0.7470
Parks and greenery	0.1264	0.2383	0.5195	0.5528	0.5919	0.5850
Abandoned equipment	0.5743	0.2980	0.7933	0.4577	0.5246	0.7118
Illegal parking	0.1760	0.1283	0.4883	0.4067	0.2527	0.5719
Drains and sewers	0.0000	0.3401	0.5999	0.5761	0.5641	0.7105
Water bodies	0.0152	0.4950	0.6711	0.6540	0.6292	0.7376
Fire safety and emergency access	0.0302	0.3364	0.6374	0.6222	0.6578	0.8198
Safety	0.2384	0.4693	0.8101	0.7452	0.7864	0.8246
Public transport infrastructure	0.1043	0.3687	0.6023	0.6968	0.6843	0.7335
Street furniture and amenities	0.0916	0.3351	0.5220	0.4458	0.5465	0.5208

Localisation performance by domain

Localisation micro F2 for each domain, using exact class matching and one-to-one box matching at IoU ≥ 0.50 on the same 11,534 shared-valid images.

Domain	YOLO-World	Gemma 4 12B	GLM 5.3 Flash	GPT-5.6 Luna	Muse Glimmer 30B	Gemini 3.8 Flash
Estate maintenance	0.0273	0.1242	0.2672	0.2212	0.1984	0.2830
Roads and footpaths	0.0249	0.1641	0.3085	0.3063	0.2712	0.2786
Animals	0.1878	0.3255	0.7247	0.7062	0.5242	0.5167
Pests	0.0047	0.1245	0.3920	0.3083	0.2864	0.4308
Parks and greenery	0.0770	0.1497	0.3696	0.3794	0.3992	0.3210
Abandoned equipment	0.3083	0.2199	0.6684	0.3516	0.3970	0.4706
Illegal parking	0.0610	0.1001	0.4563	0.3586	0.2170	0.3351
Drains and sewers	0.0000	0.1766	0.3945	0.3572	0.3542	0.4104
Water bodies	0.0063	0.1361	0.3953	0.3280	0.2960	0.3808
Fire safety and emergency access	0.0000	0.1068	0.4709	0.3710	0.3837	0.4828
Safety	0.1504	0.2668	0.6181	0.5401	0.5202	0.3711
Public transport infrastructure	0.0299	0.1208	0.2778	0.3052	0.2813	0.2934
Street furniture and amenities	0.0368	0.1631	0.3786	0.3061	0.3686	0.1811

Domain scores are point estimates; small differences do not establish a statistically significant ranking. Domain class support varies, and domain scores are not averaged to obtain the overall micro score. N/A means undefined. The full model results include per-class scores, domain precision/recall and IoU sensitivity.

Per-model valid scopes and coverage are also published. Parse failures, if present, must be counted separately and excluded from scoring, never treated as correct negatives. Scoring does not filter or rank retained detections by confidence. YOLO's inference cutoff was 0.10; VLMs had no numerical confidence cutoff. No pre-run qualification was performed for the hosted runs. GLM was restricted from a larger run; Gemini retains a known box-axis interpretation limitation. See the run provenance and evaluation protocol.

Historical 302- and 1,200-image benchmark aggregates are supplementary. Their source populations included 1 and 107 undistributed real images, respectively, and used older gold. These historical benchmarks cannot be fully reproduced from this release. They are not subsets to use for reproducing the new primary benchmark scores. No real-image assets or per-image historical records are included.

Citation and further results

The supplementary guide explains the questions, comparisons and main findings for semantic relaxation, historical benchmarks, Gemma prompt grouping, repeatability, sampling and parking prompts. It links each study to its machine-readable results.

Cite MunicipalBench using this repository URL and the exact commit SHA used in your evaluation.