An analysis using SAGAI, a vision-language workflow that interprets street images, found that just 10% of the mapped street network in Nice’s north-eastern periphery — 11.0 km — met a three-part test covering sidewalks, pedestrian-accessible entrances and vegetation. The result came from adopted visual thresholds applied to street segments.
The maps showed strong contrasts between neighborhoods and street segments, with greenery, sidewalks and entrances following different local patterns.
From sampling points to a street map
The case study covered an area of approximately 10 km² with 28,000 inhabitants in the north-eastern periphery of Nice.
Researchers generated sampling points every 20 m along the street network, using a 5 m offset. Of 5,651 points, 4,036 — 71.4% — had valid Street View coverage, while 1,615 — 28.6% — lacked imagery. The retained left- and right-facing views produced 8,072 images; front and back views were excluded.
The final spatial dataset covered 4,036 scored points across 1,505 street segments and 110.9 km of street length.
Turning images into a scorecard
SAGAI was implemented as a single Google Colab notebook with six sequential blocks covering study-area definition, image acquisition, vision-language inference and spatial aggregation with map output.
Its underlying interface provided access to 24 model checkpoints across five model families, ranging from 1B to 110B in scale.
The first three tasks used Qwen2.5-VL-32B-Instruct in standard mode. Street-frontage length estimation used Qwen2.5-VL-7B-Instruct in reasoning mode, and all runs used an A100 GPU.
At street-segment level, vegetation was averaged from ordered scores, sidewalk coverage was the proportion of images showing a sidewalk, and “constitutedness” meant pedestrian-accessible entrances per 100 m of frontage.
A segment was classified as sufficient quality only when all three adopted thresholds were met: at least four entrances per 100 m, at least 75% of observations showing a sidewalk and an average vegetation score of 5 or lower.
The map was sharply uneven
Vegetation was generally abundant on the eastern hillsides, where average scores were often below 3. It was nearly absent in several L’Ariane housing-estate segments, Drap’s faubourg main axis and La Condamine’s commercial district. Sidewalk coverage was highest in L’Ariane and relatively good in La Trinité, Drap’s faubourg and Cantaron, but poor in La Condamine.
Entrance density was highest on main streets in La Trinité and northern Drap’s faubourg, exceeding 16 pedestrian-accessible entrances per 100 m. It declined on residential hillsides and in La Condamine.
When the three measures were combined, qualifying segments represented only 10% of the street network, equivalent to 11.0 km. They were concentrated in central L’Ariane and La Trinité, with additional good streets in historic Drap and Cantaron.
A scalable diagnostic, with checks still needed
The study did not use consensus validation: each image was scored once per task.
The authors interpret contemporary vision-language models and SAGAI as capable of supporting scalable assessment of planning-relevant streetscape qualities across extensive suburban territories.
Processing speed varied by task. Object-recognition and classification indicators were computed in approximately 2.5 hours on one A100 GPU, while frontage-length estimation took approximately 35 hours.
The SAGAI workflow is openly available on GitHub and permanently archived on Zenodo.
Paper data and sources
Original title: From Street View Imagery to Street Quality Indicators: Vision Language Inference for the Suburban 15-minute City
Authors: Joan Perez, Giovanni Fusco
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text