projectsidewalk/rampnet-benchmark

RampNet Benchmark Imagery

⚠️ This benchmark is not part of the RampNet paper

It did not exist when RampNet was published. The paper's tag, v1.0-iccv2025 (August 2025), contains no benchmark/ directory at all — its evaluation was a 1,000-panorama manually labeled gold set (manual_labels/, imagery in rampnet-dataset), drawn from the same three training cities.

These 9 city splits were built eleven months later, between 2026-07-22 and 2026-07-31, as post-publication work: to test the published model on cities and imagery sources it was never trained on, and to compare it against VLM detectors.

Use this to evaluate RampNet. Do not cite it as the paper's evaluation — the ground truth, the cities, and the matching protocol all differ, so numbers measured here are not comparable with the ones in the paper.

The panoramas behind that benchmark — 9 city splits, 11.41 GB. Unlike the paper's gold set, the splits deliberately include non-US cities and a second imagery source (Mapillary as well as Google Street View), which is the whole point: the paper's own evaluation was in-domain.

It is self-contained: the records config carries the ground truth, so you can score a model against this benchmark without cloning anything. The rubrics those verdicts were made under, the per-split reviewer confidence, and the review notes stay in git at benchmark/ — read benchmark/RUBRICS.md before treating a verdict as self-explanatory, and benchmark/README.md before quoting a precision figure, because several splits carry caveats the numbers alone do not show.

Configs

config	what it is	when you want it
records	the ground truth — per-panorama metadata, model detections with their human verdict, and reviewer-marked missed ramps	scoring any model against this benchmark
native	the panoramas exactly as fetched — 4096 to 16384 px wide, depending on city and imagery source	the resolution experiment; any re-render at higher fidelity
4096x2048	the same panoramas at the model's input size	what ground-truth reviewers actually saw — gt_gallery.py renders at 4096×2048 and never native, so this is the config a second rater needs
galleries	the incremental false-positive crops shown in the operating-point A/B pass	redoing that A/B

The records config

One row per reviewed panorama, joinable to any imagery config on pano_id:

column	meaning
source	gsv or mapillary
capture_date, lat, lng, camera_heading, width, height	panorama metadata as fetched
copyright	per-record source attribution, e.g. © <contributor> / Mapillary (CC BY-SA 4.0)
detections	model detections: x_normalized, y_normalized, confidence, and verdict
missed	ramps the reviewer marked that the model did not find, each with an unsure flag
no_missed	reviewer confirmed they checked the whole panorama and found nothing missed
model_id, model_training_date, label_type	which model produced the detections

verdict is one of correct, incorrect, unsure, duplicate. unsure is an abstention and duplicate marks a second detection of an already-matched ramp — both carry the meaning the scorer gives them, so the labels mean exactly what the published precision/recall were computed against. Panoramas that were never reviewed are not included.

Labels are derived, not original. benchmark/<city>/records.jsonl and verdicts.json in git are the source of truth; this config is regenerated from them by scripts/export_benchmark.py. Verdicts get revised, imagery does not — keeping them in separate configs means a label correction never rewrites an image blob.

Configs are named by resolution, not by consumer. "Model resolution" is a relative label that becomes wrong the moment the model's input size changes, and a published path cannot be corrected later without replacing large blobs.

Note that 4096x2048 is not uniformly smaller: for splits whose native imagery is already at or near model resolution it can be larger, because it carries an extra JPEG generation. It is a fidelity artifact — it reproduces what a reviewer's eyes were on — not a compression trick.

Usage

from datasets import load_dataset

# the ground truth for one split
gt = load_dataset("projectsidewalk/rampnet-benchmark", "records", split="gainesville")
print(gt[0]["source"], gt[0]["capture_date"], gt[0]["copyright"])
print(gt[0]["detections"])      # each with x_normalized, y_normalized, confidence, verdict
print(gt[0]["missed"])          # ramps the model did not find

# the matching pixels, at the resolution reviewers saw
px = load_dataset("projectsidewalk/rampnet-benchmark", "4096x2048", split="gainesville")
by_id = {r["pano_id"]: r["image"] for r in px}
image = by_id[gt[0]["pano_id"]]

Each row of an imagery config carries:

column	meaning
pano_id	panorama id — the join key to benchmark/<city>/records.jsonl in git
city	split name
image	the image, stored as the exact source bytes, not re-encoded on write
width, height	pixel dimensions as stored
sha256	hash of those exact bytes

Verifying you have the pixels the reviewers judged

Same pano id and filename is not evidence the bytes are the ones a reviewer saw — a re-fetch from Google Street View or Mapillary can return re-stitched or re-compressed imagery. Two independent checks exist:

1. Per row. Every row carries the sha256 of its own embedded bytes. scripts/export_benchmark.py verify re-hashes every image straight out of the Parquet, so the round trip through Parquet is checked rather than assumed.
2. Against the review. benchmark/<city>/imagery_manifest.json in git records a sha256 and pixel size per panorama, pinned at the time each split was reviewed — 206 KB describing the whole archive. That is what ties these bytes to the verdicts.

python scripts/analysis/imagery_manifest.py --verify

Provenance

Field	Value
Benchmark code and labels	https://github.com/ProjectSidewalk/RampNet @ 94be2c3
Exported	2026-08-04 by scripts/export_benchmark.py
Replication ledger	docs/replication.md
How a split is built and reviewed	docs/adding_a_benchmark_city.md

Imagery sources differ by city (Google Street View and Mapillary); per-split provenance, ground truth precision/recall, and reviewer confidence are documented in benchmark/README.md.

Citation

There is no separate publication for this benchmark. Cite the paper for the pipeline and model being evaluated, and please make clear that the evaluation set is post-publication rather than the paper's own:

@inproceedings{omeara2025rampnet,
  author    = {John S. O'Meara and Jared Hwang and Zeyu Wang and Michael Saugstad and Jon E. Froehlich},
  title     = {{RampNet: A Two-Stage Pipeline for Bootstrapping Curb Ramp Detection in Streetscape Images from Open Government Metadata}},
  booktitle = {{ICCV'25 Workshop on Vision Foundation Models and Generative AI for Accessibility: Challenges and Opportunities (ICCV 2025 Workshop)}},
  year      = {2025},
  doi       = {https://doi.org/10.48550/arXiv.2508.09415},
}