Multi-platform street-level imagery — Google Street View, Baidu panorama in mainland China, Apple Look Around, and KartaView — sampled along a road network and converted into indicators you can compare: greenery, enclosure, frontage, furniture and street activity.
Sample project · a real street-view image sampled from Fifth Avenue, New York
Overview
An image is not data until something is measured from it
Anyone can look at a street view image and form an impression. Nobody can look at forty thousand of them and form a consistent one. The value is not in the imagery; it is in turning each image into a row of numbers produced by the same rule every time, so that two streets, two districts or two cities can actually be compared.
Sampling design decides what the numbers mean. Sampling every 25 metres along the network rather than once per segment stops a long arterial road counting the same as a short cul-de-sac. Four headings per point stop a single camera angle deciding how green a street looks. Capture season matters too: the same street measured in July and in January differs by more on vegetation than most streets differ from each other.
So we fix those parameters before the first image is processed, and we record them with the results, along with model versions and confidence thresholds. Every indicator traces back to the image that produced it. A number you cannot defend in a meeting is not worth delivering.
Imagery platforms
Four street-view sources, selected for the study area
Coverage, capture date, image form and redistribution rights differ by platform. We check those constraints first, then choose one source or combine several without hiding their differences.
Google Street View
The main source outside mainland China where coverage is available. Official Street View interfaces provide panoramas, fixed headings, coordinates and capture dates; use and delivery follow Google's platform terms.
Baidu Street View (China)
The practical source for projects in mainland China. We handle BD-09 coordinate conversion, sample along the road network and record the imagery dates and coverage actually returned for the study area.
Apple Maps Look Around
An additional source in cities and regions covered by Look Around. We first verify coverage and the access or usage permitted for the project, then standardise available imagery and metadata into the same sampling scheme.
KartaView
An open, contributor-built source of geotagged street-level tracks. We can collect track metadata and ordered frames, retain contributor and licence provenance, and process them through the same detection and indicator pipeline.
Real output
Output from a real sample point
One point on Fifth Avenue in New York, at 40.7536, -73.9804. Everything below came out of the same pipeline that runs on a full project; only the scale differs. The same pipeline also runs on Baidu, Apple Look Around and KartaView imagery when those are the better source for the area.
The full panorama
A 360° equirectangular image at 4096 × 2048, which is the frame every indicator is measured from. Working from the panorama rather than a single snapshot means no part of the street is outside the field of view.
4096 × 2048 · 360°
Split into four headings
The same point rendered at 0°, 90°, 180° and 270°. Indicators are computed per heading and then averaged, so a single camera direction cannot decide how green or how enclosed a street appears.
0°
90°
180°
270°
Segmented into classes
Every pixel is assigned to a class — road, sidewalk, building, vegetation, sky, person, pole, fence — and the class shares become the green view index, the enclosure ratio and the rest of the indicator set.
Input
Segmented
Ten years of the same point
Eight capture dates returned for this location between 2017 and 2026. Re-running one sampling design across vintages is what turns a snapshot into a change measurement; note the gaps, since platforms do not revisit every street every year.
2017.11
2018.08
2019.06
2020.11
2021.05
2022.06
2024.09
2026.04
KartaView as an imagery source
Alongside Google, Baidu and Apple, we also pull open street-level tracks from KartaView. Left: coverage on the map, with each blue path marking a recorded track. Right: the raw frame sequence behind one track — ordered, timestamped images that feed the same indicator pipeline.
Sample project data. Points were sampled every 25 metres with four headings each; the capture month is carried through because vegetation indicators are not comparable across seasons.
Sample project · Street view indicators aggregated by segment
Segment
Points
Green view %
Enclosure
Storefronts / 100 m
Vehicles / img
Captured
Mill Lane
48
34.2
0.68
1.2
1.8
2025-07
Harbour Road
126
11.7
0.39
0.4
7.4
2025-07
Beech Street
62
29.5
0.91
2.1
2.6
2025-08
Market Row
34
6.8
1.71
9.6
3.1
2025-08
Vale Crescent
71
41.3
0.55
0.0
1.2
2025-07
Foundry Way
58
9.4
1.48
3.8
11.2
2025-06
Delivered work
This service on a real project
Sample projects built on this service, with the numbers they produced and what each one settled.
GIS AnalysisComputer VisionMap Visualization
Street Scene Analysis
Measuring the physical character of streets at scale using semantic segmentation, object detection and colour analysis on street-level imagery.
Network analysed
310 km
Images processed
24,800
Indicators per segment
7
What it showed
Green view index across the network ranged from 4.1% to 38.9%, and the distribution followed district boundaries far more closely than the team expected. Two adjacent districts differed by more than twenty points with no change in street type.
Enclosure ratio and green view were only weakly related. Several streets scored well on greenery while feeling open and exposed, which matters because those two qualities are often treated as one in streetscape policy.
Building a comparable category census across five metro areas, and the standardisation work that made the comparison valid.
Raw records
94,200
Unique POIs
71,480
Category labels
14 → 1
What it showed
Ranked by absolute count, Metro A led by a wide margin. Ranked per capita, it placed third, and the two markets the team had considered marginal turned out to be the most densely served.
Chain share varied from 17.9% to 51.4% across markets that industry commentary had treated as broadly similar. That spread became the central finding of the research rather than a footnote.
Which streets or area, which indicators, and what the comparison is meant to settle.
Step 02
We define the data scope
We agree which platform(s) to use — Google, Baidu, Apple Look Around or KartaView — plus licence terms, sampling interval, headings and the indicator list.
Step 03
We collect and process the data
Imagery is sampled along the network, then segmented, detected and measured with fixed model versions.
Step 04
We validate the dataset
A manual sample is scored by hand and compared against the automated indicators before anything is delivered.
Step 05
We deliver the final result
Point and segment layers, the image manifest, the method note and the validation results.
Projects are priced on network length, sampling density, the number of indicators and whether imagery has to be licensed. Tell us the area and the indicators you need and you get a fixed price before collection starts.
FAQ
Questions we get asked
What is street view data?
Street view data is what you get when street-level imagery is measured rather than looked at. Images are sampled at fixed intervals along a road network and each one is converted into values — how much vegetation is in view, how enclosed the street is, how many vehicles and shopfronts are present — which are then joined back to the street geometry as an ordinary spatial dataset.
Where does the imagery come from?
We collect across multiple street-view platforms, chosen per project and always within each platform's terms: Google Street View (Street View Static API) outside mainland China; Baidu panorama inside mainland China; Apple Maps Look Around where coverage and licence allow; and open imagery platforms such as KartaView and Mapillary, which publish under Creative Commons licences. We can also process imagery you already own or capture yourself, or commercially licensed imagery bought for the project. We do not bulk-download imagery from platforms that prohibit it, and we will say at the enquiry stage if the source you have in mind cannot be used the way you are imagining.
Do you cover streets inside mainland China?
Yes, through Baidu's panorama API, which is the practical source there since Google has no coverage. The workflow is the same — sample points along the network, retrieve fixed headings, measure the same indicators — with two differences worth planning around. Coordinates need converting between WGS 84 and the BD-09 system the platform uses, which we handle and document. And historical vintages are far less consistently available than Google's, so change-over-time studies are something we scope against what the platform actually returns for your area rather than promise up front.
Can you deliver the images themselves, or only the numbers?
The indicators are always yours to keep. The imagery depends on where it came from: open-licensed and client-owned imagery can be delivered with the dataset, while imagery retrieved through a platform API is governed by that platform's caching and redistribution terms, so what you receive is the derived measurements plus a manifest identifying each source image. We confirm which case applies before work starts, in writing.
How accurate are the indicators?
We validate every project against a manual sample scored by hand. On a typical greenery index the automated and manual scores agree within about three percentage points, which is sufficient for comparing streets against each other and not sufficient for treating any single value as an absolute truth. The validation result is delivered with the data rather than described in general terms.
How often can it be refreshed?
As often as the underlying imagery is refreshed, which for platform sources is typically every one to three years in dense urban areas and less often elsewhere. Re-running the same sampling design against new imagery gives a genuine change measurement; we report which points actually have newer imagery rather than silently mixing vintages.
Can you use imagery we captured ourselves?
Yes, and it is often the better option. Dashcam footage, 360 camera runs or an existing image library can all be processed, provided positions and headings are recoverable. Your own capture removes licence restrictions on redistribution, fixes the capture date to a moment you chose, and lets you cover streets the platforms have not visited recently.
Tell us the geography, the fields and the cadence you need for street view data. You get a scoped plan, a sample and a fixed price before any work starts.
