Vehicles along a street-level survey

Find every vehicle the RAPID survey vehicle drove past in Spokane, place it on the map, crop the two best views and, optionally, grade its condition. No inventory is needed: the objects come from Mapillary’s segmentation of the survey imagery.

Source: examples/vehicle_detection_from_street.py.

import rapidtools as rt
from rapidtools.models import GenerationConfig, load

REGION = rt.BoundingBox(min_x=-117.495, min_y=47.708, max_x=-117.482, max_y=47.717)
[token_path] = rt.download_dataset('mapillary_token')
token = token_path.read_text().strip()

pipeline = rt.Pipeline([
    # 1. Vehicles from Mapillary's detections, placed by camera geometry.
    rt.MapillaryFeatureExtractor(
        classes=['vehicles'],
        access_token=token,
        region=REGION,
        start_date='2026-08-01', end_date='2026-08-31',   # one survey window
        filter_rapid_only=True,
        frame_spacing_m=3.0,
        min_observations=3,
        merge_pieces_of_neighbours=True,   # the default, named for the driveways
        reid=True,                         # repeated passes merged by appearance
    ),
    # 2. The two closest views of each vehicle, with its outline drawn.
    rt.MapillaryObjectImageExtractor(
        'output/spokane_vehicles/crops',
        access_token=token,
        max_images_per_asset=2,
        image_size='2048',
        crop_buffer='40%',
        overlay_asset_outline=True,
        outline_shape='rotated_bbox',
        outline_color='red',
        outline_width=1,
    ),
    # 3. A model confirms each object is a vehicle (needs a Google API key).
    rt.DetectionVerifier(
        load('gemini', api_key='AIza...', model_id='gemini-3.5-flash-lite'),
        min_visible_fraction=0.5,
    ),
    # 4. Condition from a vision-language model (optional; off in the script).
    rt.AssetAnalyzer(
        load('gemini', api_key='AIza...'),
        prompt='Return JSON with a "condition" key: intact, damaged, debris or unclear.',
        generation=GenerationConfig(json_mode=True, temperature=0.0),
    ),
])
vehicles = pipeline.run(rt.PhysicalAssetCollection())
vehicles.to_geojson('spokane_vehicles.geojson', ignore_properties=['image_assets'])

What to look at in the output

• localization is 'triangulated' for vehicles whose track was intersected from two or more camera positions and 'single_view' for vehicles seen without parallax. Triangulated positions are typically within a metre; single-view ones depend on the camera height and a flat road, so treat them as approximate.

• position_sigma_m, parallax_deg, n_images and position_rms_m tell you how well constrained each position is; filter on position_sigma_m for a map you can trust.

• Vehicles that were driving, anything on the line the camera itself drove, single views whose nearest sighting is beyond 30 m, and objects that nearby same-day frames should have seen but did not, are dropped before the collection is built; the log reports how many of each.

• One vehicle, one object. A car cut in two by a pole, or a track that slid onto the car behind it where a driveway is seen end-on, used to leave a second object a few metres from the real one. Such pieces are folded into the vehicle their sightings show (merge_pieces_of_neighbours), and an object never seen within 20 m is not trusted on its own: it joins the close pass it points at. With reid=True, vehicles seen on separate passes that geometry could not join are compared by appearance (CLIP on a 2048-pixel crop each) and merged when they match; only objects seen at least 100 pixels wide take part.

• The verifier removes objects whose closest crops a model says are not a motor vehicle, or are less than half visible, and writes verify_accepted, verify_label, verify_confidence and verify_visible_fraction. On the August 2026 survey of this neighbourhood it keeps roughly 450 of 650 objects, most of the rest being far or partial views.

• observations lists every sighting with its image ID, so the crops can be regenerated later without re-running detection.

• A vehicle has to appear in at least three distinct frames (min_observations=3); lower it to keep vehicles glimpsed once or twice, at the price of more false detections.

• 'vehicles' means motor vehicles. Trailers, caravans and boats are their own classes ('trailers', 'boats'), and the verifier’s question excludes them too, so a boat on a trailer is not counted as a car.

The detection stage downloads no images, only detection metadata, so the 1 km neighbourhood above runs in about six minutes, most of it reading detections; a whole city takes about an hour and a quarter. The crop stage fetches each 2048 pixel source image once and cuts every vehicle seen in it; the appearance step fetches one more per compared vehicle.