Phantom Capacity in Urban Pedestrian Networks 🏙️🚶‍♂️

An AI-Assisted Approach to Functional Walkability Using Street View Imagery and SAM 3

---

👥 Authors

Yusuf Eminoğlu ¹ ✉️, Hilmi Evren Erdin ¹ ✉️, Nurseli Şanlı ² ✉️

¹ Department of City and Regional Planning, Faculty Of Architecture, Dokuz Eylül University, Izmir
² Department of City and Regional Planning, Graduate School, Istanbul Technical University, Istanbul 📧 Contact: [yusuf.eminoglu@deu.edu.tr](mailto:yusuf.eminoglu@deu.edu.tr) | [evren.erdin@deu.edu.tr](mailto:evren.erdin@deu.edu.tr) | [sanli25@itu.edu.tr](mailto:sanli25@itu.edu.tr)

---

📖 Abstract

Traditional urban planning frequently relies on nominal cadastral dimensions to evaluate pedestrian infrastructure, generating a measurement gap that ignores microscale street-level encroachments. This study introduces the concept of "phantom capacity"—the proportion of the designated pedestrian right-of-way functionally expropriated by static and transient obstacles—to measure actual walkability.

We evaluate 63 measurement nodes across seven distinct street axes in the dense urban center of İzmir, Türkiye. By integrating Google Street View (GSV) imagery with the zero-shot panoptic segmentation capabilities of the Segment Anything Model 3 (SAM 3), we extract precise ground-plane obstacle footprints to calculate the Effective Clear Width (ECW).

The results demonstrate that approximately 35% of the surveyed pedestrian network is completely obstructed, driven primarily by vehicular parking and commercial furniture. Statistical analysis indicates that intra-morphology variance frequently exceeds inter-morphology variance (Kruskal–Wallis H = 7.474, p = 0.058), confirming that nominal street classifications fail to guarantee accessibility outcomes.

Keywords: Phantom Capacity; Urban Walkability; Segment Anything Model (SAM 3); Pedestrian Infrastructure; Google Street View (GSV).

---

📂 Repository Structure

While the full manuscript is undergoing peer review, this repository serves as an open-access platform for the methodological outputs, high-resolution visual evidence, and supplementary materials associated with the study.

sam3paper/
├── data/
│   ├── images/                               # Original Google Street View (GSV) imagery
│   └── processed/                            # SAM 3 processing outputs
│       ├── annotated_images/                 # Visual overlays of detected features
│       ├── pixel_masks/                      # Binary/categorical pixel masks for spatial calculations
│       └── sam3_masks/                       # High-fidelity masks generated by SAM 3 capabilities
│
├── src/                                      # Open-Source Methodology Core Scripts
│   ├── gsv_data_acquisition.py               # Algorithmic imagery harvesting via GSV API
│   ├── sam3_panoptic_segmentation.py         # The primary zero-shot segmentation model logic
│   ├── baseline_pixel_segmentation.py        # Semantic mapping benchmark logic
│   ├── metrics_extraction.py                 # Extrapolating ECW and physical pixel dimensions
│   └── spatial_bottleneck_analysis.py        # Root-cause inference and topological mapping 
│
└── supplementary/                            # Extended datasets and figures
    ├── figs01.png                            # Supplementary Figure 1
    ├── figs02.png                            # Supplementary Figure 2
    ├── figs03.png                            # Supplementary Figure 3
    ├── TableS1_Node_Level_Data_p1.png        # Node-Level Metrics & Measurements (Part 1)
    └── TableS1_Node_Level_Data_p2.png        # Node-Level Metrics & Measurements (Part 2)

Highlights of the Open Data & Code:

• src/: Our end-to-end framework. The provided Python scripts contain our primary SAM 3 architectural integration, enabling independent researchers to replicate the spatial geometry extraction methodology.
• data/processed/sam3_masks/: Demonstrates the cutting-edge zero-shot segmentation of urban obstacles using Meta's SAM 3 framework.
• supplementary/TableS1_...: Includes the extensive dataset (Table S1) covering exact spatial clearances and ECW (Effective Clear Width) metrics for 63 cross-sectional nodes. It was excluded from the main manuscript text due to word limit constraints, but remains pivotal for replication.

---

🔬 Methodology Brief

Our methodology hinges on an automated computer vision pipeline:

1. Acquisition: Systematic sampling of ground-level vistas via Google Street View.
2. Segmentation: Zero-shot panoptic extraction of pedestrian zones vs. urban obstacles utilizing SAM 3.
3. Quantification: Algorithmic calculation of the Effective Clear Width (ECW) versus Nominal Width.
4. Analysis: Extraction of the Phantom Capacity ratio to re-evaluate functional walkability in dense typologies.

---

📑 Citation & License

If you utilize the datasets, segmentation masks, or the phantom capacity conceptual framework in your research, please cite this repository. A formalized DOI / bibtex citation will be available upon publication.

(C) 2026 Yusuf Eminoğlu, Hilmi Evren Erdin, Nurseli Şanlı - All rights reserved.