Downloads · 30 days
0
CienciaCiudades/yolo26-urban-segmentation
yolo26-urban-segmentation is a machine learning model from CienciaCiudades. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
This repository contains a set of YOLO26 models trained for instance segmentation, specifically focused on the urban environment at the pedestrian level. The central objective is to automate the identification and del…
Downloads · 30 days
0
Access
Public
Updated Jun 30, 2026
Repo size
70.1 MB
Likes
0
Public
Click a slice to open those files.
.pt70.1 MB · 100%
From the Hugging Face model README
This repository contains a set of YOLO26 models trained for instance segmentation, specifically focused on the urban environment at the pedestrian level. The central objective is to automate the identification and delimitation of two exact categories to calculate fundamental urban metrics:
vegetation): Used for measuring the Green View Index (GVI) on sidewalks.sky): Used for measuring the Sky View Factor (SVF).The models are designed to operate optimally on street-level images (Street View), primarily capturing sidewalks, intersections, and parks.
To build the dataset, the alivestreets tool was used, enabling the automated download of Google Street View (GSV) images.
Following extensive architectural ablation studies, the YOLO "small" (s) capacity was selected as the standard. The reduced parameter count acts as a structural regularizer against overfitting, forcing the network to learn fundamental, generalizable urban features within the restricted two-class semantic space.
The following weights (.pt) are available in this repository:
mexicali_yolo26s.pt: Trained exclusively with the curated dataset from Mexicali.cdmx_yolo26s.pt: Trained exclusively with the curated dataset from Mexico City (CDMX).combined_yolo26s.pt: Trained by merging both datasets to ensure high generalization capabilities across varying geographical regions and urban topologies.The following metrics reflect the real performance of the models on the unseen Test set. The reported Precision, Recall, and mAP@50 strictly correspond to the Mask (segmentation) predictions, and the F1-Score is calculated directly from the overall Precision and Recall.
| Model | Architecture | mAP@50 | Precision | Recall | F1-Score |
|---|---|---|---|---|---|
| Mexicali | YOLO26s | 78.4% | 80.8% | 73.7% | 77.1% |
| CDMX | YOLO26s | 85.0% | 80.2% | 80.3% | 80.2% |
| Combined | YOLO26s | 82.8% | 85.5% | 75.4% | 80.1% |
In previous iterations, the models exhibited asymmetric performance due to a moderate class imbalance (approximately 1,000 sky instances vs. 4,500 vegetation instances) inherent in Street View imagery.
To mathematically resolve this in the YOLO26 architecture, the training pipeline utilizes a custom local configuration that applies a partial inverse frequency penalty (cls_pw: 0.5). This configuration scales down the optimization pressure of the majority class, successfully stabilizing the global loss gradient without overcompensating.
However, instance boundaries for vegetation in dense urban environments (like CDMX) remain highly complex due to overlapping tree canopies, narrow streets, and severe shadow casting. Resolving these dense, irregular polygon boundaries remains the primary bottleneck for achieving a higher recall score on the vegetation class.
With the successful migration to YOLO26 and local environment execution, future development will prioritize data-centric and algorithmic rigor: