docs/evaluation.md

Evaluation and reproduction

The primary benchmark scores the six released prediction files against the final annotations for all 11,534 images. Use exact class identity for the main comparison. Semantic relaxation and older experiments are explained in the supplementary guide.

What counts as a correct prediction?

Classification compares the set of predicted classes with the gold set on each image. A class present in both is a true positive (TP); an extra predicted class is a false positive (FP); a missed gold class is a false negative (FN). Repeated detections of the same class count once for classification.

Localisation compares individual boxes. A correct match requires the same class and intersection-over-union (IoU) of at least 0.50. IoU is the area shared by two boxes divided by their combined area. Maximum-cardinality matching finds the largest possible set of one-to-one matches: neither box can be used twice. Each match is a TP; unmatched predictions and gold boxes are FP and FN.

For example, suppose gold contains two litter boxes and one pothole box. A prediction with two litter boxes and no pothole has classification TP=1 and FN=1. If only one predicted litter box matches a gold box at the required IoU, localisation has TP=1, FP=1 and FN=2. Predicting the right class does not by itself localise every instance.

Metrics and populations

Precision is TP / (TP + FP) and recall is TP / (TP + FN). The benchmark emphasises recall through F2 = 5TP / (5TP + FP + 4FN). Scores are reported on a 0–1 scale. Micro metrics pool counts across classes before computing a score; macro metrics average class-level scores over the classes represented in the metric output. Domain micro scores pool only that domain's classes; averaging domain scores does not recover the overall micro score.

A valid empty prediction contributes missed targets on a positive image and no TP/FP/FN on a negative image. Correct negatives therefore do not inflate F2. The separate negative-image false-positive rate measures the fraction of clean images with at least one predicted violation. A malformed response is excluded and counted as invalid, never treated as a correct negative.

valid scores use each model's valid images. common scores use their intersection across the compared models. All six published runs have 11,534 valid responses, so the two scopes coincide here. Always report coverage when scoring another run; invalid or missing responses can change the population.

Evaluation retains every saved detection without confidence filtering, weighting or ranking. Inference retention is separate: YOLO used conf=0.10, and VLMs used no numerical confidence cutoff, although output-count caps can apply. These metrics are not detection AP/mAP or ROC-AUC. Exact localisation at IoU 0.30 and 0.40 is also supplied as sensitivity analysis.

See model provenance for recorded settings and limitations, including Gemini's box-axis interpretation.

Reproduce the published primary results

Download the full dataset snapshot, including the image files, and run these commands from its root. Install dependencies in your existing Python environment:

python -m pip install -r code/requirements.txt
python examples/load_dataset.py --root .
python examples/render_annotation.py --root . --output /tmp/municipalbench.png
python examples/score_predictions.py --root . --output /tmp/model-scores.json
python examples/reproduce_agreement.py --root . --output /tmp/municipalbench-iaa

Example	Successful result
Load	11,534 unique image IDs, matching streaming and local loads
Render	An image with final gold boxes and class IDs at the chosen output path
Score	Exact classification/localisation summaries and per-class counts match both published scopes
Agreement	All first-pass groups, estimates and confidence intervals match

Scoring uses the supplied predictions; it makes no model API calls. Agreement reproduction uses 2,000 image-bootstrap replicates and can take several minutes. Outputs are written outside the downloaded inventory. Preserve the snapshot and record its HF commit SHA when reporting results.

code/ contains the scoring and agreement implementation, with file hashes in code/source_inventory.json. The examples verify primary scoring and agreement; they do not rerun inference, recreate image generation, or replay historical studies. The supplied semantic matrix and metric code support secondary rescoring, but the scoring example above checks exact primary metrics only.